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