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
 791         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
 792                         GFP_KERNEL);
 793         if (unlikely(!pages)) {
 794                 ret |= VM_FAULT_OOM;
 795                 goto out;
 796         }
 797
 798         for (i = 0; i < HPAGE_PMD_NR; i++) {
 799                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
 800                                                __GFP_OTHER_NODE,
 801                                                vma, address, page_to_nid(page));
 802                 if (unlikely(!pages[i] ||
 803                              mem_cgroup_newpage_charge(pages[i], mm,
 804                                                        GFP_KERNEL))) {
 805                         if (pages[i])
 806                                 put_page(pages[i]);
 807                         mem_cgroup_uncharge_start();
 808                         while (--i >= 0) {
 809                                 mem_cgroup_uncharge_page(pages[i]);
 810                                 put_page(pages[i]);
 811                         }
 812                         mem_cgroup_uncharge_end();
 813                         kfree(pages);
 814                         ret |= VM_FAULT_OOM;
 815                         goto out;
 816                 }
 817         }
 818
 819         for (i = 0; i < HPAGE_PMD_NR; i++) {
 820                 copy_user_highpage(pages[i], page + i,
 821                                    haddr + PAGE_SIZE * i, vma);
 822                 __SetPageUptodate(pages[i]);
 823                 cond_resched();
 824         }
 825
 826         spin_lock(&mm->page_table_lock);
 827         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 828                 goto out_free_pages;
 829         VM_BUG_ON(!PageHead(page));
 830
 831         pmdp_clear_flush_notify(vma, haddr, pmd);
 832         /* leave pmd empty until pte is filled */
 833
 834         pgtable = pgtable_trans_huge_withdraw(mm);
 835         pmd_populate(mm, &_pmd, pgtable);
 836
 837         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 838                 pte_t *pte, entry;
 839                 entry = mk_pte(pages[i], vma->vm_page_prot);
 840                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 841                 page_add_new_anon_rmap(pages[i], vma, haddr);
 842                 pte = pte_offset_map(&_pmd, haddr);
 843                 VM_BUG_ON(!pte_none(*pte));
 844                 set_pte_at(mm, haddr, pte, entry);
 845                 pte_unmap(pte);
 846         }
 847         kfree(pages);
 848
 849         smp_wmb(); /* make pte visible before pmd */
 850         pmd_populate(mm, pmd, pgtable);
 851         page_remove_rmap(page);
 852         spin_unlock(&mm->page_table_lock);
 853
 854         ret |= VM_FAULT_WRITE;
 855         put_page(page);
 856
 857 out:
 858         return ret;
 859
 860 out_free_pages:
 861         spin_unlock(&mm->page_table_lock);
 862         mem_cgroup_uncharge_start();
 863         for (i = 0; i < HPAGE_PMD_NR; i++) {
 864                 mem_cgroup_uncharge_page(pages[i]);
 865                 put_page(pages[i]);
 866         }
 867         mem_cgroup_uncharge_end();
 868         kfree(pages);
 869         goto out;
 870 }
 871
 872 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
 873                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
 874 {
 875         int ret = 0;
 876         struct page *page, *new_page;
 877         unsigned long haddr;
 878
 879         VM_BUG_ON(!vma->anon_vma);
 880         spin_lock(&mm->page_table_lock);
 881         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 882                 goto out_unlock;
 883
 884         page = pmd_page(orig_pmd);
 885         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
 886         haddr = address & HPAGE_PMD_MASK;
 887         if (page_mapcount(page) == 1) {
 888                 pmd_t entry;
 889                 entry = pmd_mkyoung(orig_pmd);
 890                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 891                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
 892                         update_mmu_cache(vma, address, entry);
 893                 ret |= VM_FAULT_WRITE;
 894                 goto out_unlock;
 895         }
 896         get_page(page);
 897         spin_unlock(&mm->page_table_lock);
 898
 899         if (transparent_hugepage_enabled(vma) &&
 900             !transparent_hugepage_debug_cow())
 901                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 902                                               vma, haddr, numa_node_id(), 0);
 903         else
 904                 new_page = NULL;
 905
 906         if (unlikely(!new_page)) {
 907                 count_vm_event(THP_FAULT_FALLBACK);
 908                 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
 909                                                    pmd, orig_pmd, page, haddr);
 910                 if (ret & VM_FAULT_OOM)
 911                         split_huge_page(page);
 912                 put_page(page);
 913                 goto out;
 914         }
 915         count_vm_event(THP_FAULT_ALLOC);
 916
 917         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
 918                 put_page(new_page);
 919                 split_huge_page(page);
 920                 put_page(page);
 921                 ret |= VM_FAULT_OOM;
 922                 goto out;
 923         }
 924
 925         copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
 926         __SetPageUptodate(new_page);
 927
 928         spin_lock(&mm->page_table_lock);
 929         put_page(page);
 930         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
 931                 spin_unlock(&mm->page_table_lock);
 932                 mem_cgroup_uncharge_page(new_page);
 933                 put_page(new_page);
 934                 goto out;
 935         } else {
 936                 pmd_t entry;
 937                 VM_BUG_ON(!PageHead(page));
 938                 entry = mk_pmd(new_page, vma->vm_page_prot);
 939                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 940                 entry = pmd_mkhuge(entry);
 941                 pmdp_clear_flush_notify(vma, haddr, pmd);
 942                 page_add_new_anon_rmap(new_page, vma, haddr);
 943                 set_pmd_at(mm, haddr, pmd, entry);
 944                 update_mmu_cache(vma, address, entry);
 945                 page_remove_rmap(page);
 946                 put_page(page);
 947                 ret |= VM_FAULT_WRITE;
 948         }
 949 out_unlock:
 950         spin_unlock(&mm->page_table_lock);
 951 out:
 952         return ret;
 953 }
 954
 955 struct page *follow_trans_huge_pmd(struct mm_struct *mm,
 956                                    unsigned long addr,
 957                                    pmd_t *pmd,
 958                                    unsigned int flags)
 959 {
 960         struct page *page = NULL;
 961
 962         assert_spin_locked(&mm->page_table_lock);
 963
 964         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 965                 goto out;
 966
 967         page = pmd_page(*pmd);
 968         VM_BUG_ON(!PageHead(page));
 969         if (flags & FOLL_TOUCH) {
 970                 pmd_t _pmd;
 971                 /*
 972                  * We should set the dirty bit only for FOLL_WRITE but
 973                  * for now the dirty bit in the pmd is meaningless.
 974                  * And if the dirty bit will become meaningful and
 975                  * we'll only set it with FOLL_WRITE, an atomic
 976                  * set_bit will be required on the pmd to set the
 977                  * young bit, instead of the current set_pmd_at.
 978                  */
 979                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
 980                 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
 981         }
 982         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
 983         VM_BUG_ON(!PageCompound(page));
 984         if (flags & FOLL_GET)
 985                 get_page_foll(page);
 986
 987 out:
 988         return page;
 989 }
 990
 991 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
 992                  pmd_t *pmd, unsigned long addr)
 993 {
 994         int ret = 0;
 995
 996         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
 997                 struct page *page;
 998                 pgtable_t pgtable;
 999                 pgtable = pgtable_trans_huge_withdraw(tlb->mm);
1000                 page = pmd_page(*pmd);
1001                 pmd_clear(pmd);
1002                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1003                 page_remove_rmap(page);
1004                 VM_BUG_ON(page_mapcount(page) < 0);
1005                 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1006                 VM_BUG_ON(!PageHead(page));
1007                 tlb->mm->nr_ptes--;
1008                 spin_unlock(&tlb->mm->page_table_lock);
1009                 tlb_remove_page(tlb, page);
1010                 pte_free(tlb->mm, pgtable);
1011                 ret = 1;
1012         }
1013         return ret;
1014 }
1015
1016 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1017                 unsigned long addr, unsigned long end,
1018                 unsigned char *vec)
1019 {
1020         int ret = 0;
1021
1022         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1023                 /*
1024                  * All logical pages in the range are present
1025                  * if backed by a huge page.
1026                  */
1027                 spin_unlock(&vma->vm_mm->page_table_lock);
1028                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1029                 ret = 1;
1030         }
1031
1032         return ret;
1033 }
1034
1035 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1036                   unsigned long old_addr,
1037                   unsigned long new_addr, unsigned long old_end,
1038                   pmd_t *old_pmd, pmd_t *new_pmd)
1039 {
1040         int ret = 0;
1041         pmd_t pmd;
1042
1043         struct mm_struct *mm = vma->vm_mm;
1044
1045         if ((old_addr & ~HPAGE_PMD_MASK) ||
1046             (new_addr & ~HPAGE_PMD_MASK) ||
1047             old_end - old_addr < HPAGE_PMD_SIZE ||
1048             (new_vma->vm_flags & VM_NOHUGEPAGE))
1049                 goto out;
1050
1051         /*
1052          * The destination pmd shouldn't be established, free_pgtables()
1053          * should have release it.
1054          */
1055         if (WARN_ON(!pmd_none(*new_pmd))) {
1056                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1057                 goto out;
1058         }
1059
1060         ret = __pmd_trans_huge_lock(old_pmd, vma);
1061         if (ret == 1) {
1062                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1063                 VM_BUG_ON(!pmd_none(*new_pmd));
1064                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1065                 spin_unlock(&mm->page_table_lock);
1066         }
1067 out:
1068         return ret;
1069 }
1070
1071 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1072                 unsigned long addr, pgprot_t newprot)
1073 {
1074         struct mm_struct *mm = vma->vm_mm;
1075         int ret = 0;
1076
1077         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1078                 pmd_t entry;
1079                 entry = pmdp_get_and_clear(mm, addr, pmd);
1080                 entry = pmd_modify(entry, newprot);
1081                 set_pmd_at(mm, addr, pmd, entry);
1082                 spin_unlock(&vma->vm_mm->page_table_lock);
1083                 ret = 1;
1084         }
1085
1086         return ret;
1087 }
1088
1089 /*
1090  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1091  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1092  *
1093  * Note that if it returns 1, this routine returns without unlocking page
1094  * table locks. So callers must unlock them.
1095  */
1096 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1097 {
1098         spin_lock(&vma->vm_mm->page_table_lock);
1099         if (likely(pmd_trans_huge(*pmd))) {
1100                 if (unlikely(pmd_trans_splitting(*pmd))) {
1101                         spin_unlock(&vma->vm_mm->page_table_lock);
1102                         wait_split_huge_page(vma->anon_vma, pmd);
1103                         return -1;
1104                 } else {
1105                         /* Thp mapped by 'pmd' is stable, so we can
1106                          * handle it as it is. */
1107                         return 1;
1108                 }
1109         }
1110         spin_unlock(&vma->vm_mm->page_table_lock);
1111         return 0;
1112 }
1113
1114 pmd_t *page_check_address_pmd(struct page *page,
1115                               struct mm_struct *mm,
1116                               unsigned long address,
1117                               enum page_check_address_pmd_flag flag)
1118 {
1119         pgd_t *pgd;
1120         pud_t *pud;
1121         pmd_t *pmd, *ret = NULL;
1122
1123         if (address & ~HPAGE_PMD_MASK)
1124                 goto out;
1125
1126         pgd = pgd_offset(mm, address);
1127         if (!pgd_present(*pgd))
1128                 goto out;
1129
1130         pud = pud_offset(pgd, address);
1131         if (!pud_present(*pud))
1132                 goto out;
1133
1134         pmd = pmd_offset(pud, address);
1135         if (pmd_none(*pmd))
1136                 goto out;
1137         if (pmd_page(*pmd) != page)
1138                 goto out;
1139         /*
1140          * split_vma() may create temporary aliased mappings. There is
1141          * no risk as long as all huge pmd are found and have their
1142          * splitting bit set before __split_huge_page_refcount
1143          * runs. Finding the same huge pmd more than once during the
1144          * same rmap walk is not a problem.
1145          */
1146         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1147             pmd_trans_splitting(*pmd))
1148                 goto out;
1149         if (pmd_trans_huge(*pmd)) {
1150                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1151                           !pmd_trans_splitting(*pmd));
1152                 ret = pmd;
1153         }
1154 out:
1155         return ret;
1156 }
1157
1158 static int __split_huge_page_splitting(struct page *page,
1159                                        struct vm_area_struct *vma,
1160                                        unsigned long address)
1161 {
1162         struct mm_struct *mm = vma->vm_mm;
1163         pmd_t *pmd;
1164         int ret = 0;
1165
1166         spin_lock(&mm->page_table_lock);
1167         pmd = page_check_address_pmd(page, mm, address,
1168                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1169         if (pmd) {
1170                 /*
1171                  * We can't temporarily set the pmd to null in order
1172                  * to split it, the pmd must remain marked huge at all
1173                  * times or the VM won't take the pmd_trans_huge paths
1174                  * and it won't wait on the anon_vma->root->mutex to
1175                  * serialize against split_huge_page*.
1176                  */
1177                 pmdp_splitting_flush_notify(vma, address, pmd);
1178                 ret = 1;
1179         }
1180         spin_unlock(&mm->page_table_lock);
1181
1182         return ret;
1183 }
1184
1185 static void __split_huge_page_refcount(struct page *page)
1186 {
1187         int i;
1188         struct zone *zone = page_zone(page);
1189         struct lruvec *lruvec;
1190         int tail_count = 0;
1191
1192         /* prevent PageLRU to go away from under us, and freeze lru stats */
1193         spin_lock_irq(&zone->lru_lock);
1194         lruvec = mem_cgroup_page_lruvec(page, zone);
1195
1196         compound_lock(page);
1197         /* complete memcg works before add pages to LRU */
1198         mem_cgroup_split_huge_fixup(page);
1199
1200         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1201                 struct page *page_tail = page + i;
1202
1203                 /* tail_page->_mapcount cannot change */
1204                 BUG_ON(page_mapcount(page_tail) < 0);
1205                 tail_count += page_mapcount(page_tail);
1206                 /* check for overflow */
1207                 BUG_ON(tail_count < 0);
1208                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1209                 /*
1210                  * tail_page->_count is zero and not changing from
1211                  * under us. But get_page_unless_zero() may be running
1212                  * from under us on the tail_page. If we used
1213                  * atomic_set() below instead of atomic_add(), we
1214                  * would then run atomic_set() concurrently with
1215                  * get_page_unless_zero(), and atomic_set() is
1216                  * implemented in C not using locked ops. spin_unlock
1217                  * on x86 sometime uses locked ops because of PPro
1218                  * errata 66, 92, so unless somebody can guarantee
1219                  * atomic_set() here would be safe on all archs (and
1220                  * not only on x86), it's safer to use atomic_add().
1221                  */
1222                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1223                            &page_tail->_count);
1224
1225                 /* after clearing PageTail the gup refcount can be released */
1226                 smp_mb();
1227
1228                 /*
1229                  * retain hwpoison flag of the poisoned tail page:
1230                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1231                  *   by the memory-failure.
1232                  */
1233                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1234                 page_tail->flags |= (page->flags &
1235                                      ((1L << PG_referenced) |
1236                                       (1L << PG_swapbacked) |
1237                                       (1L << PG_mlocked) |
1238                                       (1L << PG_uptodate)));
1239                 page_tail->flags |= (1L << PG_dirty);
1240
1241                 /* clear PageTail before overwriting first_page */
1242                 smp_wmb();
1243
1244                 /*
1245                  * __split_huge_page_splitting() already set the
1246                  * splitting bit in all pmd that could map this
1247                  * hugepage, that will ensure no CPU can alter the
1248                  * mapcount on the head page. The mapcount is only
1249                  * accounted in the head page and it has to be
1250                  * transferred to all tail pages in the below code. So
1251                  * for this code to be safe, the split the mapcount
1252                  * can't change. But that doesn't mean userland can't
1253                  * keep changing and reading the page contents while
1254                  * we transfer the mapcount, so the pmd splitting
1255                  * status is achieved setting a reserved bit in the
1256                  * pmd, not by clearing the present bit.
1257                 */
1258                 page_tail->_mapcount = page->_mapcount;
1259
1260                 BUG_ON(page_tail->mapping);
1261                 page_tail->mapping = page->mapping;
1262
1263                 page_tail->index = page->index + i;
1264
1265                 BUG_ON(!PageAnon(page_tail));
1266                 BUG_ON(!PageUptodate(page_tail));
1267                 BUG_ON(!PageDirty(page_tail));
1268                 BUG_ON(!PageSwapBacked(page_tail));
1269
1270                 lru_add_page_tail(page, page_tail, lruvec);
1271         }
1272         atomic_sub(tail_count, &page->_count);
1273         BUG_ON(atomic_read(&page->_count) <= 0);
1274
1275         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1276         __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1277
1278         ClearPageCompound(page);
1279         compound_unlock(page);
1280         spin_unlock_irq(&zone->lru_lock);
1281
1282         for (i = 1; i < HPAGE_PMD_NR; i++) {
1283                 struct page *page_tail = page + i;
1284                 BUG_ON(page_count(page_tail) <= 0);
1285                 /*
1286                  * Tail pages may be freed if there wasn't any mapping
1287                  * like if add_to_swap() is running on a lru page that
1288                  * had its mapping zapped. And freeing these pages
1289                  * requires taking the lru_lock so we do the put_page
1290                  * of the tail pages after the split is complete.
1291                  */
1292                 put_page(page_tail);
1293         }
1294
1295         /*
1296          * Only the head page (now become a regular page) is required
1297          * to be pinned by the caller.
1298          */
1299         BUG_ON(page_count(page) <= 0);
1300 }
1301
1302 static int __split_huge_page_map(struct page *page,
1303                                  struct vm_area_struct *vma,
1304                                  unsigned long address)
1305 {
1306         struct mm_struct *mm = vma->vm_mm;
1307         pmd_t *pmd, _pmd;
1308         int ret = 0, i;
1309         pgtable_t pgtable;
1310         unsigned long haddr;
1311
1312         spin_lock(&mm->page_table_lock);
1313         pmd = page_check_address_pmd(page, mm, address,
1314                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1315         if (pmd) {
1316                 pgtable = pgtable_trans_huge_withdraw(mm);
1317                 pmd_populate(mm, &_pmd, pgtable);
1318
1319                 haddr = address;
1320                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1321                         pte_t *pte, entry;
1322                         BUG_ON(PageCompound(page+i));
1323                         entry = mk_pte(page + i, vma->vm_page_prot);
1324                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1325                         if (!pmd_write(*pmd))
1326                                 entry = pte_wrprotect(entry);
1327                         else
1328                                 BUG_ON(page_mapcount(page) != 1);
1329                         if (!pmd_young(*pmd))
1330                                 entry = pte_mkold(entry);
1331                         pte = pte_offset_map(&_pmd, haddr);
1332                         BUG_ON(!pte_none(*pte));
1333                         set_pte_at(mm, haddr, pte, entry);
1334                         pte_unmap(pte);
1335                 }
1336
1337                 smp_wmb(); /* make pte visible before pmd */
1338                 /*
1339                  * Up to this point the pmd is present and huge and
1340                  * userland has the whole access to the hugepage
1341                  * during the split (which happens in place). If we
1342                  * overwrite the pmd with the not-huge version
1343                  * pointing to the pte here (which of course we could
1344                  * if all CPUs were bug free), userland could trigger
1345                  * a small page size TLB miss on the small sized TLB
1346                  * while the hugepage TLB entry is still established
1347                  * in the huge TLB. Some CPU doesn't like that. See
1348                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1349                  * Erratum 383 on page 93. Intel should be safe but is
1350                  * also warns that it's only safe if the permission
1351                  * and cache attributes of the two entries loaded in
1352                  * the two TLB is identical (which should be the case
1353                  * here). But it is generally safer to never allow
1354                  * small and huge TLB entries for the same virtual
1355                  * address to be loaded simultaneously. So instead of
1356                  * doing "pmd_populate(); flush_tlb_range();" we first
1357                  * mark the current pmd notpresent (atomically because
1358                  * here the pmd_trans_huge and pmd_trans_splitting
1359                  * must remain set at all times on the pmd until the
1360                  * split is complete for this pmd), then we flush the
1361                  * SMP TLB and finally we write the non-huge version
1362                  * of the pmd entry with pmd_populate.
1363                  */
1364                 pmdp_invalidate(vma, address, pmd);
1365                 pmd_populate(mm, pmd, pgtable);
1366                 ret = 1;
1367         }
1368         spin_unlock(&mm->page_table_lock);
1369
1370         return ret;
1371 }
1372
1373 /* must be called with anon_vma->root->mutex hold */
1374 static void __split_huge_page(struct page *page,
1375                               struct anon_vma *anon_vma)
1376 {
1377         int mapcount, mapcount2;
1378         struct anon_vma_chain *avc;
1379
1380         BUG_ON(!PageHead(page));
1381         BUG_ON(PageTail(page));
1382
1383         mapcount = 0;
1384         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1385                 struct vm_area_struct *vma = avc->vma;
1386                 unsigned long addr = vma_address(page, vma);
1387                 BUG_ON(is_vma_temporary_stack(vma));
1388                 if (addr == -EFAULT)
1389                         continue;
1390                 mapcount += __split_huge_page_splitting(page, vma, addr);
1391         }
1392         /*
1393          * It is critical that new vmas are added to the tail of the
1394          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1395          * and establishes a child pmd before
1396          * __split_huge_page_splitting() freezes the parent pmd (so if
1397          * we fail to prevent copy_huge_pmd() from running until the
1398          * whole __split_huge_page() is complete), we will still see
1399          * the newly established pmd of the child later during the
1400          * walk, to be able to set it as pmd_trans_splitting too.
1401          */
1402         if (mapcount != page_mapcount(page))
1403                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1404                        mapcount, page_mapcount(page));
1405         BUG_ON(mapcount != page_mapcount(page));
1406
1407         __split_huge_page_refcount(page);
1408
1409         mapcount2 = 0;
1410         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1411                 struct vm_area_struct *vma = avc->vma;
1412                 unsigned long addr = vma_address(page, vma);
1413                 BUG_ON(is_vma_temporary_stack(vma));
1414                 if (addr == -EFAULT)
1415                         continue;
1416                 mapcount2 += __split_huge_page_map(page, vma, addr);
1417         }
1418         if (mapcount != mapcount2)
1419                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1420                        mapcount, mapcount2, page_mapcount(page));
1421         BUG_ON(mapcount != mapcount2);
1422 }
1423
1424 int split_huge_page(struct page *page)
1425 {
1426         struct anon_vma *anon_vma;
1427         int ret = 1;
1428
1429         BUG_ON(!PageAnon(page));
1430         anon_vma = page_lock_anon_vma(page);
1431         if (!anon_vma)
1432                 goto out;
1433         ret = 0;
1434         if (!PageCompound(page))
1435                 goto out_unlock;
1436
1437         BUG_ON(!PageSwapBacked(page));
1438         __split_huge_page(page, anon_vma);
1439         count_vm_event(THP_SPLIT);
1440
1441         BUG_ON(PageCompound(page));
1442 out_unlock:
1443         page_unlock_anon_vma(anon_vma);
1444 out:
1445         return ret;
1446 }
1447
1448 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1449
1450 int hugepage_madvise(struct vm_area_struct *vma,
1451                      unsigned long *vm_flags, int advice)
1452 {
1453         switch (advice) {
1454         case MADV_HUGEPAGE:
1455                 /*
1456                  * Be somewhat over-protective like KSM for now!
1457                  */
1458                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1459                         return -EINVAL;
1460                 *vm_flags &= ~VM_NOHUGEPAGE;
1461                 *vm_flags |= VM_HUGEPAGE;
1462                 /*
1463                  * If the vma become good for khugepaged to scan,
1464                  * register it here without waiting a page fault that
1465                  * may not happen any time soon.
1466                  */
1467                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1468                         return -ENOMEM;
1469                 break;
1470         case MADV_NOHUGEPAGE:
1471                 /*
1472                  * Be somewhat over-protective like KSM for now!
1473                  */
1474                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1475                         return -EINVAL;
1476                 *vm_flags &= ~VM_HUGEPAGE;
1477                 *vm_flags |= VM_NOHUGEPAGE;
1478                 /*
1479                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1480                  * this vma even if we leave the mm registered in khugepaged if
1481                  * it got registered before VM_NOHUGEPAGE was set.
1482                  */
1483                 break;
1484         }
1485
1486         return 0;
1487 }
1488
1489 static int __init khugepaged_slab_init(void)
1490 {
1491         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1492                                           sizeof(struct mm_slot),
1493                                           __alignof__(struct mm_slot), 0, NULL);
1494         if (!mm_slot_cache)
1495                 return -ENOMEM;
1496
1497         return 0;
1498 }
1499
1500 static void __init khugepaged_slab_free(void)
1501 {
1502         kmem_cache_destroy(mm_slot_cache);
1503         mm_slot_cache = NULL;
1504 }
1505
1506 static inline struct mm_slot *alloc_mm_slot(void)
1507 {
1508         if (!mm_slot_cache)     /* initialization failed */
1509                 return NULL;
1510         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1511 }
1512
1513 static inline void free_mm_slot(struct mm_slot *mm_slot)
1514 {
1515         kmem_cache_free(mm_slot_cache, mm_slot);
1516 }
1517
1518 static int __init mm_slots_hash_init(void)
1519 {
1520         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1521                                 GFP_KERNEL);
1522         if (!mm_slots_hash)
1523                 return -ENOMEM;
1524         return 0;
1525 }
1526
1527 #if 0
1528 static void __init mm_slots_hash_free(void)
1529 {
1530         kfree(mm_slots_hash);
1531         mm_slots_hash = NULL;
1532 }
1533 #endif
1534
1535 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1536 {
1537         struct mm_slot *mm_slot;
1538         struct hlist_head *bucket;
1539         struct hlist_node *node;
1540
1541         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1542                                 % MM_SLOTS_HASH_HEADS];
1543         hlist_for_each_entry(mm_slot, node, bucket, hash) {
1544                 if (mm == mm_slot->mm)
1545                         return mm_slot;
1546         }
1547         return NULL;
1548 }
1549
1550 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1551                                     struct mm_slot *mm_slot)
1552 {
1553         struct hlist_head *bucket;
1554
1555         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1556                                 % MM_SLOTS_HASH_HEADS];
1557         mm_slot->mm = mm;
1558         hlist_add_head(&mm_slot->hash, bucket);
1559 }
1560
1561 static inline int khugepaged_test_exit(struct mm_struct *mm)
1562 {
1563         return atomic_read(&mm->mm_users) == 0;
1564 }
1565
1566 int __khugepaged_enter(struct mm_struct *mm)
1567 {
1568         struct mm_slot *mm_slot;
1569         int wakeup;
1570
1571         mm_slot = alloc_mm_slot();
1572         if (!mm_slot)
1573                 return -ENOMEM;
1574
1575         /* __khugepaged_exit() must not run from under us */
1576         VM_BUG_ON(khugepaged_test_exit(mm));
1577         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1578                 free_mm_slot(mm_slot);
1579                 return 0;
1580         }
1581
1582         spin_lock(&khugepaged_mm_lock);
1583         insert_to_mm_slots_hash(mm, mm_slot);
1584         /*
1585          * Insert just behind the scanning cursor, to let the area settle
1586          * down a little.
1587          */
1588         wakeup = list_empty(&khugepaged_scan.mm_head);
1589         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1590         spin_unlock(&khugepaged_mm_lock);
1591
1592         atomic_inc(&mm->mm_count);
1593         if (wakeup)
1594                 wake_up_interruptible(&khugepaged_wait);
1595
1596         return 0;
1597 }
1598
1599 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1600 {
1601         unsigned long hstart, hend;
1602         if (!vma->anon_vma)
1603                 /*
1604                  * Not yet faulted in so we will register later in the
1605                  * page fault if needed.
1606                  */
1607                 return 0;
1608         if (vma->vm_ops)
1609                 /* khugepaged not yet working on file or special mappings */
1610                 return 0;
1611         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
1612         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1613         hend = vma->vm_end & HPAGE_PMD_MASK;
1614         if (hstart < hend)
1615                 return khugepaged_enter(vma);
1616         return 0;
1617 }
1618
1619 void __khugepaged_exit(struct mm_struct *mm)
1620 {
1621         struct mm_slot *mm_slot;
1622         int free = 0;
1623
1624         spin_lock(&khugepaged_mm_lock);
1625         mm_slot = get_mm_slot(mm);
1626         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1627                 hlist_del(&mm_slot->hash);
1628                 list_del(&mm_slot->mm_node);
1629                 free = 1;
1630         }
1631         spin_unlock(&khugepaged_mm_lock);
1632
1633         if (free) {
1634                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1635                 free_mm_slot(mm_slot);
1636                 mmdrop(mm);
1637         } else if (mm_slot) {
1638                 /*
1639                  * This is required to serialize against
1640                  * khugepaged_test_exit() (which is guaranteed to run
1641                  * under mmap sem read mode). Stop here (after we
1642                  * return all pagetables will be destroyed) until
1643                  * khugepaged has finished working on the pagetables
1644                  * under the mmap_sem.
1645                  */
1646                 down_write(&mm->mmap_sem);
1647                 up_write(&mm->mmap_sem);
1648         }
1649 }
1650
1651 static void release_pte_page(struct page *page)
1652 {
1653         /* 0 stands for page_is_file_cache(page) == false */
1654         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1655         unlock_page(page);
1656         putback_lru_page(page);
1657 }
1658
1659 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1660 {
1661         while (--_pte >= pte) {
1662                 pte_t pteval = *_pte;
1663                 if (!pte_none(pteval))
1664                         release_pte_page(pte_page(pteval));
1665         }
1666 }
1667
1668 static void release_all_pte_pages(pte_t *pte)
1669 {
1670         release_pte_pages(pte, pte + HPAGE_PMD_NR);
1671 }
1672
1673 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1674                                         unsigned long address,
1675                                         pte_t *pte)
1676 {
1677         struct page *page;
1678         pte_t *_pte;
1679         int referenced = 0, isolated = 0, none = 0;
1680         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1681              _pte++, address += PAGE_SIZE) {
1682                 pte_t pteval = *_pte;
1683                 if (pte_none(pteval)) {
1684                         if (++none <= khugepaged_max_ptes_none)
1685                                 continue;
1686                         else {
1687                                 release_pte_pages(pte, _pte);
1688                                 goto out;
1689                         }
1690                 }
1691                 if (!pte_present(pteval) || !pte_write(pteval)) {
1692                         release_pte_pages(pte, _pte);
1693                         goto out;
1694                 }
1695                 page = vm_normal_page(vma, address, pteval);
1696                 if (unlikely(!page)) {
1697                         release_pte_pages(pte, _pte);
1698                         goto out;
1699                 }
1700                 VM_BUG_ON(PageCompound(page));
1701                 BUG_ON(!PageAnon(page));
1702                 VM_BUG_ON(!PageSwapBacked(page));
1703
1704                 /* cannot use mapcount: can't collapse if there's a gup pin */
1705                 if (page_count(page) != 1) {
1706                         release_pte_pages(pte, _pte);
1707                         goto out;
1708                 }
1709                 /*
1710                  * We can do it before isolate_lru_page because the
1711                  * page can't be freed from under us. NOTE: PG_lock
1712                  * is needed to serialize against split_huge_page
1713                  * when invoked from the VM.
1714                  */
1715                 if (!trylock_page(page)) {
1716                         release_pte_pages(pte, _pte);
1717                         goto out;
1718                 }
1719                 /*
1720                  * Isolate the page to avoid collapsing an hugepage
1721                  * currently in use by the VM.
1722                  */
1723                 if (isolate_lru_page(page)) {
1724                         unlock_page(page);
1725                         release_pte_pages(pte, _pte);
1726                         goto out;
1727                 }
1728                 /* 0 stands for page_is_file_cache(page) == false */
1729                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1730                 VM_BUG_ON(!PageLocked(page));
1731                 VM_BUG_ON(PageLRU(page));
1732
1733                 /* If there is no mapped pte young don't collapse the page */
1734                 if (pte_young(pteval) || PageReferenced(page) ||
1735                     mmu_notifier_test_young(vma->vm_mm, address))
1736                         referenced = 1;
1737         }
1738         if (unlikely(!referenced))
1739                 release_all_pte_pages(pte);
1740         else
1741                 isolated = 1;
1742 out:
1743         return isolated;
1744 }
1745
1746 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1747                                       struct vm_area_struct *vma,
1748                                       unsigned long address,
1749                                       spinlock_t *ptl)
1750 {
1751         pte_t *_pte;
1752         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1753                 pte_t pteval = *_pte;
1754                 struct page *src_page;
1755
1756                 if (pte_none(pteval)) {
1757                         clear_user_highpage(page, address);
1758                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1759                 } else {
1760                         src_page = pte_page(pteval);
1761                         copy_user_highpage(page, src_page, address, vma);
1762                         VM_BUG_ON(page_mapcount(src_page) != 1);
1763                         release_pte_page(src_page);
1764                         /*
1765                          * ptl mostly unnecessary, but preempt has to
1766                          * be disabled to update the per-cpu stats
1767                          * inside page_remove_rmap().
1768                          */
1769                         spin_lock(ptl);
1770                         /*
1771                          * paravirt calls inside pte_clear here are
1772                          * superfluous.
1773                          */
1774                         pte_clear(vma->vm_mm, address, _pte);
1775                         page_remove_rmap(src_page);
1776                         spin_unlock(ptl);
1777                         free_page_and_swap_cache(src_page);
1778                 }
1779
1780                 address += PAGE_SIZE;
1781                 page++;
1782         }
1783 }
1784
1785 static void khugepaged_alloc_sleep(void)
1786 {
1787         wait_event_freezable_timeout(khugepaged_wait, false,
1788                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
1789 }
1790
1791 #ifdef CONFIG_NUMA
1792 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
1793 {
1794         if (IS_ERR(*hpage)) {
1795                 if (!*wait)
1796                         return false;
1797
1798                 *wait = false;
1799                 khugepaged_alloc_sleep();
1800         } else if (*hpage) {
1801                 put_page(*hpage);
1802                 *hpage = NULL;
1803         }
1804
1805         return true;
1806 }
1807
1808 static struct page
1809 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
1810                        struct vm_area_struct *vma, unsigned long address,
1811                        int node)
1812 {
1813         VM_BUG_ON(*hpage);
1814         /*
1815          * Allocate the page while the vma is still valid and under
1816          * the mmap_sem read mode so there is no memory allocation
1817          * later when we take the mmap_sem in write mode. This is more
1818          * friendly behavior (OTOH it may actually hide bugs) to
1819          * filesystems in userland with daemons allocating memory in
1820          * the userland I/O paths.  Allocating memory with the
1821          * mmap_sem in read mode is good idea also to allow greater
1822          * scalability.
1823          */
1824         *hpage  = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
1825                                       node, __GFP_OTHER_NODE);
1826
1827         /*
1828          * After allocating the hugepage, release the mmap_sem read lock in
1829          * preparation for taking it in write mode.
1830          */
1831         up_read(&mm->mmap_sem);
1832         if (unlikely(!*hpage)) {
1833                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1834                 *hpage = ERR_PTR(-ENOMEM);
1835                 return NULL;
1836         }
1837
1838         count_vm_event(THP_COLLAPSE_ALLOC);
1839         return *hpage;
1840 }
1841 #else
1842 static struct page *khugepaged_alloc_hugepage(bool *wait)
1843 {
1844         struct page *hpage;
1845
1846         do {
1847                 hpage = alloc_hugepage(khugepaged_defrag());
1848                 if (!hpage) {
1849                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1850                         if (!*wait)
1851                                 return NULL;
1852
1853                         *wait = false;
1854                         khugepaged_alloc_sleep();
1855                 } else
1856                         count_vm_event(THP_COLLAPSE_ALLOC);
1857         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
1858
1859         return hpage;
1860 }
1861
1862 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
1863 {
1864         if (!*hpage)
1865                 *hpage = khugepaged_alloc_hugepage(wait);
1866
1867         if (unlikely(!*hpage))
1868                 return false;
1869
1870         return true;
1871 }
1872
1873 static struct page
1874 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
1875                        struct vm_area_struct *vma, unsigned long address,
1876                        int node)
1877 {
1878         up_read(&mm->mmap_sem);
1879         VM_BUG_ON(!*hpage);
1880         return  *hpage;
1881 }
1882 #endif
1883
1884 static void collapse_huge_page(struct mm_struct *mm,
1885                                    unsigned long address,
1886                                    struct page **hpage,
1887                                    struct vm_area_struct *vma,
1888                                    int node)
1889 {
1890         pgd_t *pgd;
1891         pud_t *pud;
1892         pmd_t *pmd, _pmd;
1893         pte_t *pte;
1894         pgtable_t pgtable;
1895         struct page *new_page;
1896         spinlock_t *ptl;
1897         int isolated;
1898         unsigned long hstart, hend;
1899
1900         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1901
1902         /* release the mmap_sem read lock. */
1903         new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
1904         if (!new_page)
1905                 return;
1906
1907         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
1908                 return;
1909
1910         /*
1911          * Prevent all access to pagetables with the exception of
1912          * gup_fast later hanlded by the ptep_clear_flush and the VM
1913          * handled by the anon_vma lock + PG_lock.
1914          */
1915         down_write(&mm->mmap_sem);
1916         if (unlikely(khugepaged_test_exit(mm)))
1917                 goto out;
1918
1919         vma = find_vma(mm, address);
1920         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1921         hend = vma->vm_end & HPAGE_PMD_MASK;
1922         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
1923                 goto out;
1924
1925         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
1926             (vma->vm_flags & VM_NOHUGEPAGE))
1927                 goto out;
1928
1929         if (!vma->anon_vma || vma->vm_ops)
1930                 goto out;
1931         if (is_vma_temporary_stack(vma))
1932                 goto out;
1933         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
1934
1935         pgd = pgd_offset(mm, address);
1936         if (!pgd_present(*pgd))
1937                 goto out;
1938
1939         pud = pud_offset(pgd, address);
1940         if (!pud_present(*pud))
1941                 goto out;
1942
1943         pmd = pmd_offset(pud, address);
1944         /* pmd can't go away or become huge under us */
1945         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1946                 goto out;
1947
1948         anon_vma_lock(vma->anon_vma);
1949
1950         pte = pte_offset_map(pmd, address);
1951         ptl = pte_lockptr(mm, pmd);
1952
1953         spin_lock(&mm->page_table_lock); /* probably unnecessary */
1954         /*
1955          * After this gup_fast can't run anymore. This also removes
1956          * any huge TLB entry from the CPU so we won't allow
1957          * huge and small TLB entries for the same virtual address
1958          * to avoid the risk of CPU bugs in that area.
1959          */
1960         _pmd = pmdp_clear_flush_notify(vma, address, pmd);
1961         spin_unlock(&mm->page_table_lock);
1962
1963         spin_lock(ptl);
1964         isolated = __collapse_huge_page_isolate(vma, address, pte);
1965         spin_unlock(ptl);
1966
1967         if (unlikely(!isolated)) {
1968                 pte_unmap(pte);
1969                 spin_lock(&mm->page_table_lock);
1970                 BUG_ON(!pmd_none(*pmd));
1971                 set_pmd_at(mm, address, pmd, _pmd);
1972                 spin_unlock(&mm->page_table_lock);
1973                 anon_vma_unlock(vma->anon_vma);
1974                 goto out;
1975         }
1976
1977         /*
1978          * All pages are isolated and locked so anon_vma rmap
1979          * can't run anymore.
1980          */
1981         anon_vma_unlock(vma->anon_vma);
1982
1983         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
1984         pte_unmap(pte);
1985         __SetPageUptodate(new_page);
1986         pgtable = pmd_pgtable(_pmd);
1987
1988         _pmd = mk_pmd(new_page, vma->vm_page_prot);
1989         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1990         _pmd = pmd_mkhuge(_pmd);
1991
1992         /*
1993          * spin_lock() below is not the equivalent of smp_wmb(), so
1994          * this is needed to avoid the copy_huge_page writes to become
1995          * visible after the set_pmd_at() write.
1996          */
1997         smp_wmb();
1998
1999         spin_lock(&mm->page_table_lock);
2000         BUG_ON(!pmd_none(*pmd));
2001         page_add_new_anon_rmap(new_page, vma, address);
2002         set_pmd_at(mm, address, pmd, _pmd);
2003         update_mmu_cache(vma, address, _pmd);
2004         pgtable_trans_huge_deposit(mm, pgtable);
2005         spin_unlock(&mm->page_table_lock);
2006
2007         *hpage = NULL;
2008
2009         khugepaged_pages_collapsed++;
2010 out_up_write:
2011         up_write(&mm->mmap_sem);
2012         return;
2013
2014 out:
2015         mem_cgroup_uncharge_page(new_page);
2016         goto out_up_write;
2017 }
2018
2019 static int khugepaged_scan_pmd(struct mm_struct *mm,
2020                                struct vm_area_struct *vma,
2021                                unsigned long address,
2022                                struct page **hpage)
2023 {
2024         pgd_t *pgd;
2025         pud_t *pud;
2026         pmd_t *pmd;
2027         pte_t *pte, *_pte;
2028         int ret = 0, referenced = 0, none = 0;
2029         struct page *page;
2030         unsigned long _address;
2031         spinlock_t *ptl;
2032         int node = -1;
2033
2034         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2035
2036         pgd = pgd_offset(mm, address);
2037         if (!pgd_present(*pgd))
2038                 goto out;
2039
2040         pud = pud_offset(pgd, address);
2041         if (!pud_present(*pud))
2042                 goto out;
2043
2044         pmd = pmd_offset(pud, address);
2045         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2046                 goto out;
2047
2048         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2049         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2050              _pte++, _address += PAGE_SIZE) {
2051                 pte_t pteval = *_pte;
2052                 if (pte_none(pteval)) {
2053                         if (++none <= khugepaged_max_ptes_none)
2054                                 continue;
2055                         else
2056                                 goto out_unmap;
2057                 }
2058                 if (!pte_present(pteval) || !pte_write(pteval))
2059                         goto out_unmap;
2060                 page = vm_normal_page(vma, _address, pteval);
2061                 if (unlikely(!page))
2062                         goto out_unmap;
2063                 /*
2064                  * Chose the node of the first page. This could
2065                  * be more sophisticated and look at more pages,
2066                  * but isn't for now.
2067                  */
2068                 if (node == -1)
2069                         node = page_to_nid(page);
2070                 VM_BUG_ON(PageCompound(page));
2071                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2072                         goto out_unmap;
2073                 /* cannot use mapcount: can't collapse if there's a gup pin */
2074                 if (page_count(page) != 1)
2075                         goto out_unmap;
2076                 if (pte_young(pteval) || PageReferenced(page) ||
2077                     mmu_notifier_test_young(vma->vm_mm, address))
2078                         referenced = 1;
2079         }
2080         if (referenced)
2081                 ret = 1;
2082 out_unmap:
2083         pte_unmap_unlock(pte, ptl);
2084         if (ret)
2085                 /* collapse_huge_page will return with the mmap_sem released */
2086                 collapse_huge_page(mm, address, hpage, vma, node);
2087 out:
2088         return ret;
2089 }
2090
2091 static void collect_mm_slot(struct mm_slot *mm_slot)
2092 {
2093         struct mm_struct *mm = mm_slot->mm;
2094
2095         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2096
2097         if (khugepaged_test_exit(mm)) {
2098                 /* free mm_slot */
2099                 hlist_del(&mm_slot->hash);
2100                 list_del(&mm_slot->mm_node);
2101
2102                 /*
2103                  * Not strictly needed because the mm exited already.
2104                  *
2105                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2106                  */
2107
2108                 /* khugepaged_mm_lock actually not necessary for the below */
2109                 free_mm_slot(mm_slot);
2110                 mmdrop(mm);
2111         }
2112 }
2113
2114 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2115                                             struct page **hpage)
2116         __releases(&khugepaged_mm_lock)
2117         __acquires(&khugepaged_mm_lock)
2118 {
2119         struct mm_slot *mm_slot;
2120         struct mm_struct *mm;
2121         struct vm_area_struct *vma;
2122         int progress = 0;
2123
2124         VM_BUG_ON(!pages);
2125         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2126
2127         if (khugepaged_scan.mm_slot)
2128                 mm_slot = khugepaged_scan.mm_slot;
2129         else {
2130                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2131                                      struct mm_slot, mm_node);
2132                 khugepaged_scan.address = 0;
2133                 khugepaged_scan.mm_slot = mm_slot;
2134         }
2135         spin_unlock(&khugepaged_mm_lock);
2136
2137         mm = mm_slot->mm;
2138         down_read(&mm->mmap_sem);
2139         if (unlikely(khugepaged_test_exit(mm)))
2140                 vma = NULL;
2141         else
2142                 vma = find_vma(mm, khugepaged_scan.address);
2143
2144         progress++;
2145         for (; vma; vma = vma->vm_next) {
2146                 unsigned long hstart, hend;
2147
2148                 cond_resched();
2149                 if (unlikely(khugepaged_test_exit(mm))) {
2150                         progress++;
2151                         break;
2152                 }
2153
2154                 if ((!(vma->vm_flags & VM_HUGEPAGE) &&
2155                      !khugepaged_always()) ||
2156                     (vma->vm_flags & VM_NOHUGEPAGE)) {
2157                 skip:
2158                         progress++;
2159                         continue;
2160                 }
2161                 if (!vma->anon_vma || vma->vm_ops)
2162                         goto skip;
2163                 if (is_vma_temporary_stack(vma))
2164                         goto skip;
2165                 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2166
2167                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2168                 hend = vma->vm_end & HPAGE_PMD_MASK;
2169                 if (hstart >= hend)
2170                         goto skip;
2171                 if (khugepaged_scan.address > hend)
2172                         goto skip;
2173                 if (khugepaged_scan.address < hstart)
2174                         khugepaged_scan.address = hstart;
2175                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2176
2177                 while (khugepaged_scan.address < hend) {
2178                         int ret;
2179                         cond_resched();
2180                         if (unlikely(khugepaged_test_exit(mm)))
2181                                 goto breakouterloop;
2182
2183                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2184                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2185                                   hend);
2186                         ret = khugepaged_scan_pmd(mm, vma,
2187                                                   khugepaged_scan.address,
2188                                                   hpage);
2189                         /* move to next address */
2190                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2191                         progress += HPAGE_PMD_NR;
2192                         if (ret)
2193                                 /* we released mmap_sem so break loop */
2194                                 goto breakouterloop_mmap_sem;
2195                         if (progress >= pages)
2196                                 goto breakouterloop;
2197                 }
2198         }
2199 breakouterloop:
2200         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2201 breakouterloop_mmap_sem:
2202
2203         spin_lock(&khugepaged_mm_lock);
2204         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2205         /*
2206          * Release the current mm_slot if this mm is about to die, or
2207          * if we scanned all vmas of this mm.
2208          */
2209         if (khugepaged_test_exit(mm) || !vma) {
2210                 /*
2211                  * Make sure that if mm_users is reaching zero while
2212                  * khugepaged runs here, khugepaged_exit will find
2213                  * mm_slot not pointing to the exiting mm.
2214                  */
2215                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2216                         khugepaged_scan.mm_slot = list_entry(
2217                                 mm_slot->mm_node.next,
2218                                 struct mm_slot, mm_node);
2219                         khugepaged_scan.address = 0;
2220                 } else {
2221                         khugepaged_scan.mm_slot = NULL;
2222                         khugepaged_full_scans++;
2223                 }
2224
2225                 collect_mm_slot(mm_slot);
2226         }
2227
2228         return progress;
2229 }
2230
2231 static int khugepaged_has_work(void)
2232 {
2233         return !list_empty(&khugepaged_scan.mm_head) &&
2234                 khugepaged_enabled();
2235 }
2236
2237 static int khugepaged_wait_event(void)
2238 {
2239         return !list_empty(&khugepaged_scan.mm_head) ||
2240                 kthread_should_stop();
2241 }
2242
2243 static void khugepaged_do_scan(void)
2244 {
2245         struct page *hpage = NULL;
2246         unsigned int progress = 0, pass_through_head = 0;
2247         unsigned int pages = khugepaged_pages_to_scan;
2248         bool wait = true;
2249
2250         barrier(); /* write khugepaged_pages_to_scan to local stack */
2251
2252         while (progress < pages) {
2253                 if (!khugepaged_prealloc_page(&hpage, &wait))
2254                         break;
2255
2256                 cond_resched();
2257
2258                 if (unlikely(kthread_should_stop() || freezing(current)))
2259                         break;
2260
2261                 spin_lock(&khugepaged_mm_lock);
2262                 if (!khugepaged_scan.mm_slot)
2263                         pass_through_head++;
2264                 if (khugepaged_has_work() &&
2265                     pass_through_head < 2)
2266                         progress += khugepaged_scan_mm_slot(pages - progress,
2267                                                             &hpage);
2268                 else
2269                         progress = pages;
2270                 spin_unlock(&khugepaged_mm_lock);
2271         }
2272
2273         if (!IS_ERR_OR_NULL(hpage))
2274                 put_page(hpage);
2275 }
2276
2277 static void khugepaged_wait_work(void)
2278 {
2279         try_to_freeze();
2280
2281         if (khugepaged_has_work()) {
2282                 if (!khugepaged_scan_sleep_millisecs)
2283                         return;
2284
2285                 wait_event_freezable_timeout(khugepaged_wait,
2286                                              kthread_should_stop(),
2287                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2288                 return;
2289         }
2290
2291         if (khugepaged_enabled())
2292                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2293 }
2294
2295 static int khugepaged(void *none)
2296 {
2297         struct mm_slot *mm_slot;
2298
2299         set_freezable();
2300         set_user_nice(current, 19);
2301
2302         while (!kthread_should_stop()) {
2303                 khugepaged_do_scan();
2304                 khugepaged_wait_work();
2305         }
2306
2307         spin_lock(&khugepaged_mm_lock);
2308         mm_slot = khugepaged_scan.mm_slot;
2309         khugepaged_scan.mm_slot = NULL;
2310         if (mm_slot)
2311                 collect_mm_slot(mm_slot);
2312         spin_unlock(&khugepaged_mm_lock);
2313         return 0;
2314 }
2315
2316 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2317 {
2318         struct page *page;
2319
2320         spin_lock(&mm->page_table_lock);
2321         if (unlikely(!pmd_trans_huge(*pmd))) {
2322                 spin_unlock(&mm->page_table_lock);
2323                 return;
2324         }
2325         page = pmd_page(*pmd);
2326         VM_BUG_ON(!page_count(page));
2327         get_page(page);
2328         spin_unlock(&mm->page_table_lock);
2329
2330         split_huge_page(page);
2331
2332         put_page(page);
2333         BUG_ON(pmd_trans_huge(*pmd));
2334 }
2335
2336 static void split_huge_page_address(struct mm_struct *mm,
2337                                     unsigned long address)
2338 {
2339         pgd_t *pgd;
2340         pud_t *pud;
2341         pmd_t *pmd;
2342
2343         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2344
2345         pgd = pgd_offset(mm, address);
2346         if (!pgd_present(*pgd))
2347                 return;
2348
2349         pud = pud_offset(pgd, address);
2350         if (!pud_present(*pud))
2351                 return;
2352
2353         pmd = pmd_offset(pud, address);
2354         if (!pmd_present(*pmd))
2355                 return;
2356         /*
2357          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2358          * materialize from under us.
2359          */
2360         split_huge_page_pmd(mm, pmd);
2361 }
2362
2363 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2364                              unsigned long start,
2365                              unsigned long end,
2366                              long adjust_next)
2367 {
2368         /*
2369          * If the new start address isn't hpage aligned and it could
2370          * previously contain an hugepage: check if we need to split
2371          * an huge pmd.
2372          */
2373         if (start & ~HPAGE_PMD_MASK &&
2374             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2375             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2376                 split_huge_page_address(vma->vm_mm, start);
2377
2378         /*
2379          * If the new end address isn't hpage aligned and it could
2380          * previously contain an hugepage: check if we need to split
2381          * an huge pmd.
2382          */
2383         if (end & ~HPAGE_PMD_MASK &&
2384             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2385             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2386                 split_huge_page_address(vma->vm_mm, end);
2387
2388         /*
2389          * If we're also updating the vma->vm_next->vm_start, if the new
2390          * vm_next->vm_start isn't page aligned and it could previously
2391          * contain an hugepage: check if we need to split an huge pmd.
2392          */
2393         if (adjust_next > 0) {
2394                 struct vm_area_struct *next = vma->vm_next;
2395                 unsigned long nstart = next->vm_start;
2396                 nstart += adjust_next << PAGE_SHIFT;
2397                 if (nstart & ~HPAGE_PMD_MASK &&
2398                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2399                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2400                         split_huge_page_address(next->vm_mm, nstart);
2401         }
2402 }