mm/huge_memory.c

   1 /*
   2  *  Copyright (C) 2009  Red Hat, Inc.
   3  *
   4  *  This work is licensed under the terms of the GNU GPL, version 2. See
   5  *  the COPYING file in the top-level directory.
   6  */
   7
   8 #include <linux/mm.h>
   9 #include <linux/sched.h>
  10 #include <linux/highmem.h>
  11 #include <linux/hugetlb.h>
  12 #include <linux/mmu_notifier.h>
  13 #include <linux/rmap.h>
  14 #include <linux/swap.h>
  15 #include <linux/mm_inline.h>
  16 #include <linux/kthread.h>
  17 #include <linux/khugepaged.h>
  18 #include <linux/freezer.h>
  19 #include <linux/mman.h>
  20 #include <asm/tlb.h>
  21 #include <asm/pgalloc.h>
  22 #include "internal.h"
  23
  24 /*
  25  * By default transparent hugepage support is enabled for all mappings
  26  * and khugepaged scans all mappings. Defrag is only invoked by
  27  * khugepaged hugepage allocations and by page faults inside
  28  * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
  29  * allocations.
  30  */
  31 unsigned long transparent_hugepage_flags __read_mostly =
  32 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  33         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  34 #endif
  35 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  36         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  37 #endif
  38         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  39         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
  40
  41 /* default scan 8*512 pte (or vmas) every 30 second */
  42 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  43 static unsigned int khugepaged_pages_collapsed;
  44 static unsigned int khugepaged_full_scans;
  45 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  46 /* during fragmentation poll the hugepage allocator once every minute */
  47 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  48 static struct task_struct *khugepaged_thread __read_mostly;
  49 static DEFINE_MUTEX(khugepaged_mutex);
  50 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  51 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  52 /*
  53  * default collapse hugepages if there is at least one pte mapped like
  54  * it would have happened if the vma was large enough during page
  55  * fault.
  56  */
  57 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  58
  59 static int khugepaged(void *none);
  60 static int mm_slots_hash_init(void);
  61 static int khugepaged_slab_init(void);
  62 static void khugepaged_slab_free(void);
  63
  64 #define MM_SLOTS_HASH_HEADS 1024
  65 static struct hlist_head *mm_slots_hash __read_mostly;
  66 static struct kmem_cache *mm_slot_cache __read_mostly;
  67
  68 /**
  69  * struct mm_slot - hash lookup from mm to mm_slot
  70  * @hash: hash collision list
  71  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  72  * @mm: the mm that this information is valid for
  73  */
  74 struct mm_slot {
  75         struct hlist_node hash;
  76         struct list_head mm_node;
  77         struct mm_struct *mm;
  78 };
  79
  80 /**
  81  * struct khugepaged_scan - cursor for scanning
  82  * @mm_head: the head of the mm list to scan
  83  * @mm_slot: the current mm_slot we are scanning
  84  * @address: the next address inside that to be scanned
  85  *
  86  * There is only the one khugepaged_scan instance of this cursor structure.
  87  */
  88 struct khugepaged_scan {
  89         struct list_head mm_head;
  90         struct mm_slot *mm_slot;
  91         unsigned long address;
  92 };
  93 static struct khugepaged_scan khugepaged_scan = {
  94         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
  95 };
  96
  97
  98 static int set_recommended_min_free_kbytes(void)
  99 {
 100         struct zone *zone;
 101         int nr_zones = 0;
 102         unsigned long recommended_min;
 103         extern int min_free_kbytes;
 104
 105         if (!khugepaged_enabled())
 106                 return 0;
 107
 108         for_each_populated_zone(zone)
 109                 nr_zones++;
 110
 111         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
 112         recommended_min = pageblock_nr_pages * nr_zones * 2;
 113
 114         /*
 115          * Make sure that on average at least two pageblocks are almost free
 116          * of another type, one for a migratetype to fall back to and a
 117          * second to avoid subsequent fallbacks of other types There are 3
 118          * MIGRATE_TYPES we care about.
 119          */
 120         recommended_min += pageblock_nr_pages * nr_zones *
 121                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 122
 123         /* don't ever allow to reserve more than 5% of the lowmem */
 124         recommended_min = min(recommended_min,
 125                               (unsigned long) nr_free_buffer_pages() / 20);
 126         recommended_min <<= (PAGE_SHIFT-10);
 127
 128         if (recommended_min > min_free_kbytes)
 129                 min_free_kbytes = recommended_min;
 130         setup_per_zone_wmarks();
 131         return 0;
 132 }
 133 late_initcall(set_recommended_min_free_kbytes);
 134
 135 static int start_khugepaged(void)
 136 {
 137         int err = 0;
 138         if (khugepaged_enabled()) {
 139                 if (!khugepaged_thread)
 140                         khugepaged_thread = kthread_run(khugepaged, NULL,
 141                                                         "khugepaged");
 142                 if (unlikely(IS_ERR(khugepaged_thread))) {
 143                         printk(KERN_ERR
 144                                "khugepaged: kthread_run(khugepaged) failed\n");
 145                         err = PTR_ERR(khugepaged_thread);
 146                         khugepaged_thread = NULL;
 147                 }
 148
 149                 if (!list_empty(&khugepaged_scan.mm_head))
 150                         wake_up_interruptible(&khugepaged_wait);
 151
 152                 set_recommended_min_free_kbytes();
 153         } else if (khugepaged_thread) {
 154                 kthread_stop(khugepaged_thread);
 155                 khugepaged_thread = NULL;
 156         }
 157
 158         return err;
 159 }
 160
 161 #ifdef CONFIG_SYSFS
 162
 163 static ssize_t double_flag_show(struct kobject *kobj,
 164                                 struct kobj_attribute *attr, char *buf,
 165                                 enum transparent_hugepage_flag enabled,
 166                                 enum transparent_hugepage_flag req_madv)
 167 {
 168         if (test_bit(enabled, &transparent_hugepage_flags)) {
 169                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 170                 return sprintf(buf, "[always] madvise never\n");
 171         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 172                 return sprintf(buf, "always [madvise] never\n");
 173         else
 174                 return sprintf(buf, "always madvise [never]\n");
 175 }
 176 static ssize_t double_flag_store(struct kobject *kobj,
 177                                  struct kobj_attribute *attr,
 178                                  const char *buf, size_t count,
 179                                  enum transparent_hugepage_flag enabled,
 180                                  enum transparent_hugepage_flag req_madv)
 181 {
 182         if (!memcmp("always", buf,
 183                     min(sizeof("always")-1, count))) {
 184                 set_bit(enabled, &transparent_hugepage_flags);
 185                 clear_bit(req_madv, &transparent_hugepage_flags);
 186         } else if (!memcmp("madvise", buf,
 187                            min(sizeof("madvise")-1, count))) {
 188                 clear_bit(enabled, &transparent_hugepage_flags);
 189                 set_bit(req_madv, &transparent_hugepage_flags);
 190         } else if (!memcmp("never", buf,
 191                            min(sizeof("never")-1, count))) {
 192                 clear_bit(enabled, &transparent_hugepage_flags);
 193                 clear_bit(req_madv, &transparent_hugepage_flags);
 194         } else
 195                 return -EINVAL;
 196
 197         return count;
 198 }
 199
 200 static ssize_t enabled_show(struct kobject *kobj,
 201                             struct kobj_attribute *attr, char *buf)
 202 {
 203         return double_flag_show(kobj, attr, buf,
 204                                 TRANSPARENT_HUGEPAGE_FLAG,
 205                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 206 }
 207 static ssize_t enabled_store(struct kobject *kobj,
 208                              struct kobj_attribute *attr,
 209                              const char *buf, size_t count)
 210 {
 211         ssize_t ret;
 212
 213         ret = double_flag_store(kobj, attr, buf, count,
 214                                 TRANSPARENT_HUGEPAGE_FLAG,
 215                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 216
 217         if (ret > 0) {
 218                 int err;
 219
 220                 mutex_lock(&khugepaged_mutex);
 221                 err = start_khugepaged();
 222                 mutex_unlock(&khugepaged_mutex);
 223
 224                 if (err)
 225                         ret = err;
 226         }
 227
 228         return ret;
 229 }
 230 static struct kobj_attribute enabled_attr =
 231         __ATTR(enabled, 0644, enabled_show, enabled_store);
 232
 233 static ssize_t single_flag_show(struct kobject *kobj,
 234                                 struct kobj_attribute *attr, char *buf,
 235                                 enum transparent_hugepage_flag flag)
 236 {
 237         return sprintf(buf, "%d\n",
 238                        !!test_bit(flag, &transparent_hugepage_flags));
 239 }
 240
 241 static ssize_t single_flag_store(struct kobject *kobj,
 242                                  struct kobj_attribute *attr,
 243                                  const char *buf, size_t count,
 244                                  enum transparent_hugepage_flag flag)
 245 {
 246         unsigned long value;
 247         int ret;
 248
 249         ret = kstrtoul(buf, 10, &value);
 250         if (ret < 0)
 251                 return ret;
 252         if (value > 1)
 253                 return -EINVAL;
 254
 255         if (value)
 256                 set_bit(flag, &transparent_hugepage_flags);
 257         else
 258                 clear_bit(flag, &transparent_hugepage_flags);
 259
 260         return count;
 261 }
 262
 263 /*
 264  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 265  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 266  * memory just to allocate one more hugepage.
 267  */
 268 static ssize_t defrag_show(struct kobject *kobj,
 269                            struct kobj_attribute *attr, char *buf)
 270 {
 271         return double_flag_show(kobj, attr, buf,
 272                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 273                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 274 }
 275 static ssize_t defrag_store(struct kobject *kobj,
 276                             struct kobj_attribute *attr,
 277                             const char *buf, size_t count)
 278 {
 279         return double_flag_store(kobj, attr, buf, count,
 280                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 281                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 282 }
 283 static struct kobj_attribute defrag_attr =
 284         __ATTR(defrag, 0644, defrag_show, defrag_store);
 285
 286 #ifdef CONFIG_DEBUG_VM
 287 static ssize_t debug_cow_show(struct kobject *kobj,
 288                                 struct kobj_attribute *attr, char *buf)
 289 {
 290         return single_flag_show(kobj, attr, buf,
 291                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 292 }
 293 static ssize_t debug_cow_store(struct kobject *kobj,
 294                                struct kobj_attribute *attr,
 295                                const char *buf, size_t count)
 296 {
 297         return single_flag_store(kobj, attr, buf, count,
 298                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 299 }
 300 static struct kobj_attribute debug_cow_attr =
 301         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 302 #endif /* CONFIG_DEBUG_VM */
 303
 304 static struct attribute *hugepage_attr[] = {
 305         &enabled_attr.attr,
 306         &defrag_attr.attr,
 307 #ifdef CONFIG_DEBUG_VM
 308         &debug_cow_attr.attr,
 309 #endif
 310         NULL,
 311 };
 312
 313 static struct attribute_group hugepage_attr_group = {
 314         .attrs = hugepage_attr,
 315 };
 316
 317 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 318                                          struct kobj_attribute *attr,
 319                                          char *buf)
 320 {
 321         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 322 }
 323
 324 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 325                                           struct kobj_attribute *attr,
 326                                           const char *buf, size_t count)
 327 {
 328         unsigned long msecs;
 329         int err;
 330
 331         err = strict_strtoul(buf, 10, &msecs);
 332         if (err || msecs > UINT_MAX)
 333                 return -EINVAL;
 334
 335         khugepaged_scan_sleep_millisecs = msecs;
 336         wake_up_interruptible(&khugepaged_wait);
 337
 338         return count;
 339 }
 340 static struct kobj_attribute scan_sleep_millisecs_attr =
 341         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 342                scan_sleep_millisecs_store);
 343
 344 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 345                                           struct kobj_attribute *attr,
 346                                           char *buf)
 347 {
 348         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 349 }
 350
 351 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 352                                            struct kobj_attribute *attr,
 353                                            const char *buf, size_t count)
 354 {
 355         unsigned long msecs;
 356         int err;
 357
 358         err = strict_strtoul(buf, 10, &msecs);
 359         if (err || msecs > UINT_MAX)
 360                 return -EINVAL;
 361
 362         khugepaged_alloc_sleep_millisecs = msecs;
 363         wake_up_interruptible(&khugepaged_wait);
 364
 365         return count;
 366 }
 367 static struct kobj_attribute alloc_sleep_millisecs_attr =
 368         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 369                alloc_sleep_millisecs_store);
 370
 371 static ssize_t pages_to_scan_show(struct kobject *kobj,
 372                                   struct kobj_attribute *attr,
 373                                   char *buf)
 374 {
 375         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 376 }
 377 static ssize_t pages_to_scan_store(struct kobject *kobj,
 378                                    struct kobj_attribute *attr,
 379                                    const char *buf, size_t count)
 380 {
 381         int err;
 382         unsigned long pages;
 383
 384         err = strict_strtoul(buf, 10, &pages);
 385         if (err || !pages || pages > UINT_MAX)
 386                 return -EINVAL;
 387
 388         khugepaged_pages_to_scan = pages;
 389
 390         return count;
 391 }
 392 static struct kobj_attribute pages_to_scan_attr =
 393         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 394                pages_to_scan_store);
 395
 396 static ssize_t pages_collapsed_show(struct kobject *kobj,
 397                                     struct kobj_attribute *attr,
 398                                     char *buf)
 399 {
 400         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 401 }
 402 static struct kobj_attribute pages_collapsed_attr =
 403         __ATTR_RO(pages_collapsed);
 404
 405 static ssize_t full_scans_show(struct kobject *kobj,
 406                                struct kobj_attribute *attr,
 407                                char *buf)
 408 {
 409         return sprintf(buf, "%u\n", khugepaged_full_scans);
 410 }
 411 static struct kobj_attribute full_scans_attr =
 412         __ATTR_RO(full_scans);
 413
 414 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 415                                       struct kobj_attribute *attr, char *buf)
 416 {
 417         return single_flag_show(kobj, attr, buf,
 418                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 419 }
 420 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 421                                        struct kobj_attribute *attr,
 422                                        const char *buf, size_t count)
 423 {
 424         return single_flag_store(kobj, attr, buf, count,
 425                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 426 }
 427 static struct kobj_attribute khugepaged_defrag_attr =
 428         __ATTR(defrag, 0644, khugepaged_defrag_show,
 429                khugepaged_defrag_store);
 430
 431 /*
 432  * max_ptes_none controls if khugepaged should collapse hugepages over
 433  * any unmapped ptes in turn potentially increasing the memory
 434  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 435  * reduce the available free memory in the system as it
 436  * runs. Increasing max_ptes_none will instead potentially reduce the
 437  * free memory in the system during the khugepaged scan.
 438  */
 439 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 440                                              struct kobj_attribute *attr,
 441                                              char *buf)
 442 {
 443         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 444 }
 445 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 446                                               struct kobj_attribute *attr,
 447                                               const char *buf, size_t count)
 448 {
 449         int err;
 450         unsigned long max_ptes_none;
 451
 452         err = strict_strtoul(buf, 10, &max_ptes_none);
 453         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 454                 return -EINVAL;
 455
 456         khugepaged_max_ptes_none = max_ptes_none;
 457
 458         return count;
 459 }
 460 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 461         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 462                khugepaged_max_ptes_none_store);
 463
 464 static struct attribute *khugepaged_attr[] = {
 465         &khugepaged_defrag_attr.attr,
 466         &khugepaged_max_ptes_none_attr.attr,
 467         &pages_to_scan_attr.attr,
 468         &pages_collapsed_attr.attr,
 469         &full_scans_attr.attr,
 470         &scan_sleep_millisecs_attr.attr,
 471         &alloc_sleep_millisecs_attr.attr,
 472         NULL,
 473 };
 474
 475 static struct attribute_group khugepaged_attr_group = {
 476         .attrs = khugepaged_attr,
 477         .name = "khugepaged",
 478 };
 479
 480 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 481 {
 482         int err;
 483
 484         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 485         if (unlikely(!*hugepage_kobj)) {
 486                 printk(KERN_ERR "hugepage: failed kobject create\n");
 487                 return -ENOMEM;
 488         }
 489
 490         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 491         if (err) {
 492                 printk(KERN_ERR "hugepage: failed register hugeage group\n");
 493                 goto delete_obj;
 494         }
 495
 496         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 497         if (err) {
 498                 printk(KERN_ERR "hugepage: failed register hugeage group\n");
 499                 goto remove_hp_group;
 500         }
 501
 502         return 0;
 503
 504 remove_hp_group:
 505         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 506 delete_obj:
 507         kobject_put(*hugepage_kobj);
 508         return err;
 509 }
 510
 511 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 512 {
 513         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 514         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 515         kobject_put(hugepage_kobj);
 516 }
 517 #else
 518 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 519 {
 520         return 0;
 521 }
 522
 523 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 524 {
 525 }
 526 #endif /* CONFIG_SYSFS */
 527
 528 static int __init hugepage_init(void)
 529 {
 530         int err;
 531         struct kobject *hugepage_kobj;
 532
 533         if (!has_transparent_hugepage()) {
 534                 transparent_hugepage_flags = 0;
 535                 return -EINVAL;
 536         }
 537
 538         err = hugepage_init_sysfs(&hugepage_kobj);
 539         if (err)
 540                 return err;
 541
 542         err = khugepaged_slab_init();
 543         if (err)
 544                 goto out;
 545
 546         err = mm_slots_hash_init();
 547         if (err) {
 548                 khugepaged_slab_free();
 549                 goto out;
 550         }
 551
 552         /*
 553          * By default disable transparent hugepages on smaller systems,
 554          * where the extra memory used could hurt more than TLB overhead
 555          * is likely to save.  The admin can still enable it through /sys.
 556          */
 557         if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
 558                 transparent_hugepage_flags = 0;
 559
 560         start_khugepaged();
 561
 562         return 0;
 563 out:
 564         hugepage_exit_sysfs(hugepage_kobj);
 565         return err;
 566 }
 567 module_init(hugepage_init)
 568
 569 static int __init setup_transparent_hugepage(char *str)
 570 {
 571         int ret = 0;
 572         if (!str)
 573                 goto out;
 574         if (!strcmp(str, "always")) {
 575                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 576                         &transparent_hugepage_flags);
 577                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 578                           &transparent_hugepage_flags);
 579                 ret = 1;
 580         } else if (!strcmp(str, "madvise")) {
 581                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 582                           &transparent_hugepage_flags);
 583                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 584                         &transparent_hugepage_flags);
 585                 ret = 1;
 586         } else if (!strcmp(str, "never")) {
 587                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 588                           &transparent_hugepage_flags);
 589                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 590                           &transparent_hugepage_flags);
 591                 ret = 1;
 592         }
 593 out:
 594         if (!ret)
 595                 printk(KERN_WARNING
 596                        "transparent_hugepage= cannot parse, ignored\n");
 597         return ret;
 598 }
 599 __setup("transparent_hugepage=", setup_transparent_hugepage);
 600
 601 static inline pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 602 {
 603         if (likely(vma->vm_flags & VM_WRITE))
 604                 pmd = pmd_mkwrite(pmd);
 605         return pmd;
 606 }
 607
 608 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 609                                         struct vm_area_struct *vma,
 610                                         unsigned long haddr, pmd_t *pmd,
 611                                         struct page *page)
 612 {
 613         pgtable_t pgtable;
 614
 615         VM_BUG_ON(!PageCompound(page));
 616         pgtable = pte_alloc_one(mm, haddr);
 617         if (unlikely(!pgtable))
 618                 return VM_FAULT_OOM;
 619
 620         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 621         __SetPageUptodate(page);
 622
 623         spin_lock(&mm->page_table_lock);
 624         if (unlikely(!pmd_none(*pmd))) {
 625                 spin_unlock(&mm->page_table_lock);
 626                 mem_cgroup_uncharge_page(page);
 627                 put_page(page);
 628                 pte_free(mm, pgtable);
 629         } else {
 630                 pmd_t entry;
 631                 entry = mk_pmd(page, vma->vm_page_prot);
 632                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 633                 entry = pmd_mkhuge(entry);
 634                 /*
 635                  * The spinlocking to take the lru_lock inside
 636                  * page_add_new_anon_rmap() acts as a full memory
 637                  * barrier to be sure clear_huge_page writes become
 638                  * visible after the set_pmd_at() write.
 639                  */
 640                 page_add_new_anon_rmap(page, vma, haddr);
 641                 set_pmd_at(mm, haddr, pmd, entry);
 642                 pgtable_trans_huge_deposit(mm, pgtable);
 643                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 644                 mm->nr_ptes++;
 645                 spin_unlock(&mm->page_table_lock);
 646         }
 647
 648         return 0;
 649 }
 650
 651 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 652 {
 653         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 654 }
 655
 656 static inline struct page *alloc_hugepage_vma(int defrag,
 657                                               struct vm_area_struct *vma,
 658                                               unsigned long haddr, int nd,
 659                                               gfp_t extra_gfp)
 660 {
 661         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
 662                                HPAGE_PMD_ORDER, vma, haddr, nd);
 663 }
 664
 665 #ifndef CONFIG_NUMA
 666 static inline struct page *alloc_hugepage(int defrag)
 667 {
 668         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
 669                            HPAGE_PMD_ORDER);
 670 }
 671 #endif
 672
 673 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 674                                unsigned long address, pmd_t *pmd,
 675                                unsigned int flags)
 676 {
 677         struct page *page;
 678         unsigned long haddr = address & HPAGE_PMD_MASK;
 679         pte_t *pte;
 680
 681         if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
 682                 if (unlikely(anon_vma_prepare(vma)))
 683                         return VM_FAULT_OOM;
 684                 if (unlikely(khugepaged_enter(vma)))
 685                         return VM_FAULT_OOM;
 686                 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 687                                           vma, haddr, numa_node_id(), 0);
 688                 if (unlikely(!page)) {
 689                         count_vm_event(THP_FAULT_FALLBACK);
 690                         goto out;
 691                 }
 692                 count_vm_event(THP_FAULT_ALLOC);
 693                 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
 694                         put_page(page);
 695                         goto out;
 696                 }
 697                 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd,
 698                                                           page))) {
 699                         mem_cgroup_uncharge_page(page);
 700                         put_page(page);
 701                         goto out;
 702                 }
 703
 704                 return 0;
 705         }
 706 out:
 707         /*
 708          * Use __pte_alloc instead of pte_alloc_map, because we can't
 709          * run pte_offset_map on the pmd, if an huge pmd could
 710          * materialize from under us from a different thread.
 711          */
 712         if (unlikely(__pte_alloc(mm, vma, pmd, address)))
 713                 return VM_FAULT_OOM;
 714         /* if an huge pmd materialized from under us just retry later */
 715         if (unlikely(pmd_trans_huge(*pmd)))
 716                 return 0;
 717         /*
 718          * A regular pmd is established and it can't morph into a huge pmd
 719          * from under us anymore at this point because we hold the mmap_sem
 720          * read mode and khugepaged takes it in write mode. So now it's
 721          * safe to run pte_offset_map().
 722          */
 723         pte = pte_offset_map(pmd, address);
 724         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
 725 }
 726
 727 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 728                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 729                   struct vm_area_struct *vma)
 730 {
 731         struct page *src_page;
 732         pmd_t pmd;
 733         pgtable_t pgtable;
 734         int ret;
 735
 736         ret = -ENOMEM;
 737         pgtable = pte_alloc_one(dst_mm, addr);
 738         if (unlikely(!pgtable))
 739                 goto out;
 740
 741         spin_lock(&dst_mm->page_table_lock);
 742         spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
 743
 744         ret = -EAGAIN;
 745         pmd = *src_pmd;
 746         if (unlikely(!pmd_trans_huge(pmd))) {
 747                 pte_free(dst_mm, pgtable);
 748                 goto out_unlock;
 749         }
 750         if (unlikely(pmd_trans_splitting(pmd))) {
 751                 /* split huge page running from under us */
 752                 spin_unlock(&src_mm->page_table_lock);
 753                 spin_unlock(&dst_mm->page_table_lock);
 754                 pte_free(dst_mm, pgtable);
 755
 756                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 757                 goto out;
 758         }
 759         src_page = pmd_page(pmd);
 760         VM_BUG_ON(!PageHead(src_page));
 761         get_page(src_page);
 762         page_dup_rmap(src_page);
 763         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 764
 765         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 766         pmd = pmd_mkold(pmd_wrprotect(pmd));
 767         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 768         pgtable_trans_huge_deposit(dst_mm, pgtable);
 769         dst_mm->nr_ptes++;
 770
 771         ret = 0;
 772 out_unlock:
 773         spin_unlock(&src_mm->page_table_lock);
 774         spin_unlock(&dst_mm->page_table_lock);
 775 out:
 776         return ret;
 777 }
 778
 779 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
 780                                         struct vm_area_struct *vma,
 781                                         unsigned long address,
 782                                         pmd_t *pmd, pmd_t orig_pmd,
 783                                         struct page *page,
 784                                         unsigned long haddr)
 785 {
 786         pgtable_t pgtable;
 787         pmd_t _pmd;
 788         int ret = 0, i;
 789         struct page **pages;
 790         unsigned long mmun_start;       /* For mmu_notifiers */
 791         unsigned long mmun_end;         /* For mmu_notifiers */
 792
 793         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
 794                         GFP_KERNEL);
 795         if (unlikely(!pages)) {
 796                 ret |= VM_FAULT_OOM;
 797                 goto out;
 798         }
 799
 800         for (i = 0; i < HPAGE_PMD_NR; i++) {
 801                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
 802                                                __GFP_OTHER_NODE,
 803                                                vma, address, page_to_nid(page));
 804                 if (unlikely(!pages[i] ||
 805                              mem_cgroup_newpage_charge(pages[i], mm,
 806                                                        GFP_KERNEL))) {
 807                         if (pages[i])
 808                                 put_page(pages[i]);
 809                         mem_cgroup_uncharge_start();
 810                         while (--i >= 0) {
 811                                 mem_cgroup_uncharge_page(pages[i]);
 812                                 put_page(pages[i]);
 813                         }
 814                         mem_cgroup_uncharge_end();
 815                         kfree(pages);
 816                         ret |= VM_FAULT_OOM;
 817                         goto out;
 818                 }
 819         }
 820
 821         for (i = 0; i < HPAGE_PMD_NR; i++) {
 822                 copy_user_highpage(pages[i], page + i,
 823                                    haddr + PAGE_SIZE * i, vma);
 824                 __SetPageUptodate(pages[i]);
 825                 cond_resched();
 826         }
 827
 828         mmun_start = haddr;
 829         mmun_end   = haddr + HPAGE_PMD_SIZE;
 830         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
 831
 832         spin_lock(&mm->page_table_lock);
 833         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 834                 goto out_free_pages;
 835         VM_BUG_ON(!PageHead(page));
 836
 837         pmdp_clear_flush(vma, haddr, pmd);
 838         /* leave pmd empty until pte is filled */
 839
 840         pgtable = pgtable_trans_huge_withdraw(mm);
 841         pmd_populate(mm, &_pmd, pgtable);
 842
 843         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 844                 pte_t *pte, entry;
 845                 entry = mk_pte(pages[i], vma->vm_page_prot);
 846                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 847                 page_add_new_anon_rmap(pages[i], vma, haddr);
 848                 pte = pte_offset_map(&_pmd, haddr);
 849                 VM_BUG_ON(!pte_none(*pte));
 850                 set_pte_at(mm, haddr, pte, entry);
 851                 pte_unmap(pte);
 852         }
 853         kfree(pages);
 854
 855         smp_wmb(); /* make pte visible before pmd */
 856         pmd_populate(mm, pmd, pgtable);
 857         page_remove_rmap(page);
 858         spin_unlock(&mm->page_table_lock);
 859
 860         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
 861
 862         ret |= VM_FAULT_WRITE;
 863         put_page(page);
 864
 865 out:
 866         return ret;
 867
 868 out_free_pages:
 869         spin_unlock(&mm->page_table_lock);
 870         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
 871         mem_cgroup_uncharge_start();
 872         for (i = 0; i < HPAGE_PMD_NR; i++) {
 873                 mem_cgroup_uncharge_page(pages[i]);
 874                 put_page(pages[i]);
 875         }
 876         mem_cgroup_uncharge_end();
 877         kfree(pages);
 878         goto out;
 879 }
 880
 881 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
 882                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
 883 {
 884         int ret = 0;
 885         struct page *page, *new_page;
 886         unsigned long haddr;
 887         unsigned long mmun_start;       /* For mmu_notifiers */
 888         unsigned long mmun_end;         /* For mmu_notifiers */
 889
 890         VM_BUG_ON(!vma->anon_vma);
 891         spin_lock(&mm->page_table_lock);
 892         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 893                 goto out_unlock;
 894
 895         page = pmd_page(orig_pmd);
 896         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
 897         haddr = address & HPAGE_PMD_MASK;
 898         if (page_mapcount(page) == 1) {
 899                 pmd_t entry;
 900                 entry = pmd_mkyoung(orig_pmd);
 901                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 902                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
 903                         update_mmu_cache(vma, address, pmd);
 904                 ret |= VM_FAULT_WRITE;
 905                 goto out_unlock;
 906         }
 907         get_page(page);
 908         spin_unlock(&mm->page_table_lock);
 909
 910         if (transparent_hugepage_enabled(vma) &&
 911             !transparent_hugepage_debug_cow())
 912                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 913                                               vma, haddr, numa_node_id(), 0);
 914         else
 915                 new_page = NULL;
 916
 917         if (unlikely(!new_page)) {
 918                 count_vm_event(THP_FAULT_FALLBACK);
 919                 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
 920                                                    pmd, orig_pmd, page, haddr);
 921                 if (ret & VM_FAULT_OOM)
 922                         split_huge_page(page);
 923                 put_page(page);
 924                 goto out;
 925         }
 926         count_vm_event(THP_FAULT_ALLOC);
 927
 928         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
 929                 put_page(new_page);
 930                 split_huge_page(page);
 931                 put_page(page);
 932                 ret |= VM_FAULT_OOM;
 933                 goto out;
 934         }
 935
 936         copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
 937         __SetPageUptodate(new_page);
 938
 939         mmun_start = haddr;
 940         mmun_end   = haddr + HPAGE_PMD_SIZE;
 941         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
 942
 943         spin_lock(&mm->page_table_lock);
 944         put_page(page);
 945         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
 946                 spin_unlock(&mm->page_table_lock);
 947                 mem_cgroup_uncharge_page(new_page);
 948                 put_page(new_page);
 949                 goto out_mn;
 950         } else {
 951                 pmd_t entry;
 952                 VM_BUG_ON(!PageHead(page));
 953                 entry = mk_pmd(new_page, vma->vm_page_prot);
 954                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 955                 entry = pmd_mkhuge(entry);
 956                 pmdp_clear_flush(vma, haddr, pmd);
 957                 page_add_new_anon_rmap(new_page, vma, haddr);
 958                 set_pmd_at(mm, haddr, pmd, entry);
 959                 update_mmu_cache(vma, address, pmd);
 960                 page_remove_rmap(page);
 961                 put_page(page);
 962                 ret |= VM_FAULT_WRITE;
 963         }
 964         spin_unlock(&mm->page_table_lock);
 965 out_mn:
 966         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
 967 out:
 968         return ret;
 969 out_unlock:
 970         spin_unlock(&mm->page_table_lock);
 971         return ret;
 972 }
 973
 974 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
 975                                    unsigned long addr,
 976                                    pmd_t *pmd,
 977                                    unsigned int flags)
 978 {
 979         struct mm_struct *mm = vma->vm_mm;
 980         struct page *page = NULL;
 981
 982         assert_spin_locked(&mm->page_table_lock);
 983
 984         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 985                 goto out;
 986
 987         page = pmd_page(*pmd);
 988         VM_BUG_ON(!PageHead(page));
 989         if (flags & FOLL_TOUCH) {
 990                 pmd_t _pmd;
 991                 /*
 992                  * We should set the dirty bit only for FOLL_WRITE but
 993                  * for now the dirty bit in the pmd is meaningless.
 994                  * And if the dirty bit will become meaningful and
 995                  * we'll only set it with FOLL_WRITE, an atomic
 996                  * set_bit will be required on the pmd to set the
 997                  * young bit, instead of the current set_pmd_at.
 998                  */
 999                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1000                 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
1001         }
1002         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1003                 if (page->mapping && trylock_page(page)) {
1004                         lru_add_drain();
1005                         if (page->mapping)
1006                                 mlock_vma_page(page);
1007                         unlock_page(page);
1008                 }
1009         }
1010         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1011         VM_BUG_ON(!PageCompound(page));
1012         if (flags & FOLL_GET)
1013                 get_page_foll(page);
1014
1015 out:
1016         return page;
1017 }
1018
1019 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1020                  pmd_t *pmd, unsigned long addr)
1021 {
1022         int ret = 0;
1023
1024         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1025                 struct page *page;
1026                 pgtable_t pgtable;
1027                 pgtable = pgtable_trans_huge_withdraw(tlb->mm);
1028                 page = pmd_page(*pmd);
1029                 pmd_clear(pmd);
1030                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1031                 page_remove_rmap(page);
1032                 VM_BUG_ON(page_mapcount(page) < 0);
1033                 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1034                 VM_BUG_ON(!PageHead(page));
1035                 tlb->mm->nr_ptes--;
1036                 spin_unlock(&tlb->mm->page_table_lock);
1037                 tlb_remove_page(tlb, page);
1038                 pte_free(tlb->mm, pgtable);
1039                 ret = 1;
1040         }
1041         return ret;
1042 }
1043
1044 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1045                 unsigned long addr, unsigned long end,
1046                 unsigned char *vec)
1047 {
1048         int ret = 0;
1049
1050         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1051                 /*
1052                  * All logical pages in the range are present
1053                  * if backed by a huge page.
1054                  */
1055                 spin_unlock(&vma->vm_mm->page_table_lock);
1056                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1057                 ret = 1;
1058         }
1059
1060         return ret;
1061 }
1062
1063 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1064                   unsigned long old_addr,
1065                   unsigned long new_addr, unsigned long old_end,
1066                   pmd_t *old_pmd, pmd_t *new_pmd)
1067 {
1068         int ret = 0;
1069         pmd_t pmd;
1070
1071         struct mm_struct *mm = vma->vm_mm;
1072
1073         if ((old_addr & ~HPAGE_PMD_MASK) ||
1074             (new_addr & ~HPAGE_PMD_MASK) ||
1075             old_end - old_addr < HPAGE_PMD_SIZE ||
1076             (new_vma->vm_flags & VM_NOHUGEPAGE))
1077                 goto out;
1078
1079         /*
1080          * The destination pmd shouldn't be established, free_pgtables()
1081          * should have release it.
1082          */
1083         if (WARN_ON(!pmd_none(*new_pmd))) {
1084                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1085                 goto out;
1086         }
1087
1088         ret = __pmd_trans_huge_lock(old_pmd, vma);
1089         if (ret == 1) {
1090                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1091                 VM_BUG_ON(!pmd_none(*new_pmd));
1092                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1093                 spin_unlock(&mm->page_table_lock);
1094         }
1095 out:
1096         return ret;
1097 }
1098
1099 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1100                 unsigned long addr, pgprot_t newprot)
1101 {
1102         struct mm_struct *mm = vma->vm_mm;
1103         int ret = 0;
1104
1105         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1106                 pmd_t entry;
1107                 entry = pmdp_get_and_clear(mm, addr, pmd);
1108                 entry = pmd_modify(entry, newprot);
1109                 set_pmd_at(mm, addr, pmd, entry);
1110                 spin_unlock(&vma->vm_mm->page_table_lock);
1111                 ret = 1;
1112         }
1113
1114         return ret;
1115 }
1116
1117 /*
1118  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1119  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1120  *
1121  * Note that if it returns 1, this routine returns without unlocking page
1122  * table locks. So callers must unlock them.
1123  */
1124 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1125 {
1126         spin_lock(&vma->vm_mm->page_table_lock);
1127         if (likely(pmd_trans_huge(*pmd))) {
1128                 if (unlikely(pmd_trans_splitting(*pmd))) {
1129                         spin_unlock(&vma->vm_mm->page_table_lock);
1130                         wait_split_huge_page(vma->anon_vma, pmd);
1131                         return -1;
1132                 } else {
1133                         /* Thp mapped by 'pmd' is stable, so we can
1134                          * handle it as it is. */
1135                         return 1;
1136                 }
1137         }
1138         spin_unlock(&vma->vm_mm->page_table_lock);
1139         return 0;
1140 }
1141
1142 pmd_t *page_check_address_pmd(struct page *page,
1143                               struct mm_struct *mm,
1144                               unsigned long address,
1145                               enum page_check_address_pmd_flag flag)
1146 {
1147         pgd_t *pgd;
1148         pud_t *pud;
1149         pmd_t *pmd, *ret = NULL;
1150
1151         if (address & ~HPAGE_PMD_MASK)
1152                 goto out;
1153
1154         pgd = pgd_offset(mm, address);
1155         if (!pgd_present(*pgd))
1156                 goto out;
1157
1158         pud = pud_offset(pgd, address);
1159         if (!pud_present(*pud))
1160                 goto out;
1161
1162         pmd = pmd_offset(pud, address);
1163         if (pmd_none(*pmd))
1164                 goto out;
1165         if (pmd_page(*pmd) != page)
1166                 goto out;
1167         /*
1168          * split_vma() may create temporary aliased mappings. There is
1169          * no risk as long as all huge pmd are found and have their
1170          * splitting bit set before __split_huge_page_refcount
1171          * runs. Finding the same huge pmd more than once during the
1172          * same rmap walk is not a problem.
1173          */
1174         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1175             pmd_trans_splitting(*pmd))
1176                 goto out;
1177         if (pmd_trans_huge(*pmd)) {
1178                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1179                           !pmd_trans_splitting(*pmd));
1180                 ret = pmd;
1181         }
1182 out:
1183         return ret;
1184 }
1185
1186 static int __split_huge_page_splitting(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;
1192         int ret = 0;
1193         /* For mmu_notifiers */
1194         const unsigned long mmun_start = address;
1195         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1196
1197         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1198         spin_lock(&mm->page_table_lock);
1199         pmd = page_check_address_pmd(page, mm, address,
1200                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1201         if (pmd) {
1202                 /*
1203                  * We can't temporarily set the pmd to null in order
1204                  * to split it, the pmd must remain marked huge at all
1205                  * times or the VM won't take the pmd_trans_huge paths
1206                  * and it won't wait on the anon_vma->root->mutex to
1207                  * serialize against split_huge_page*.
1208                  */
1209                 pmdp_splitting_flush(vma, address, pmd);
1210                 ret = 1;
1211         }
1212         spin_unlock(&mm->page_table_lock);
1213         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1214
1215         return ret;
1216 }
1217
1218 static void __split_huge_page_refcount(struct page *page)
1219 {
1220         int i;
1221         struct zone *zone = page_zone(page);
1222         struct lruvec *lruvec;
1223         int tail_count = 0;
1224
1225         /* prevent PageLRU to go away from under us, and freeze lru stats */
1226         spin_lock_irq(&zone->lru_lock);
1227         lruvec = mem_cgroup_page_lruvec(page, zone);
1228
1229         compound_lock(page);
1230         /* complete memcg works before add pages to LRU */
1231         mem_cgroup_split_huge_fixup(page);
1232
1233         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1234                 struct page *page_tail = page + i;
1235
1236                 /* tail_page->_mapcount cannot change */
1237                 BUG_ON(page_mapcount(page_tail) < 0);
1238                 tail_count += page_mapcount(page_tail);
1239                 /* check for overflow */
1240                 BUG_ON(tail_count < 0);
1241                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1242                 /*
1243                  * tail_page->_count is zero and not changing from
1244                  * under us. But get_page_unless_zero() may be running
1245                  * from under us on the tail_page. If we used
1246                  * atomic_set() below instead of atomic_add(), we
1247                  * would then run atomic_set() concurrently with
1248                  * get_page_unless_zero(), and atomic_set() is
1249                  * implemented in C not using locked ops. spin_unlock
1250                  * on x86 sometime uses locked ops because of PPro
1251                  * errata 66, 92, so unless somebody can guarantee
1252                  * atomic_set() here would be safe on all archs (and
1253                  * not only on x86), it's safer to use atomic_add().
1254                  */
1255                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1256                            &page_tail->_count);
1257
1258                 /* after clearing PageTail the gup refcount can be released */
1259                 smp_mb();
1260
1261                 /*
1262                  * retain hwpoison flag of the poisoned tail page:
1263                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1264                  *   by the memory-failure.
1265                  */
1266                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1267                 page_tail->flags |= (page->flags &
1268                                      ((1L << PG_referenced) |
1269                                       (1L << PG_swapbacked) |
1270                                       (1L << PG_mlocked) |
1271                                       (1L << PG_uptodate)));
1272                 page_tail->flags |= (1L << PG_dirty);
1273
1274                 /* clear PageTail before overwriting first_page */
1275                 smp_wmb();
1276
1277                 /*
1278                  * __split_huge_page_splitting() already set the
1279                  * splitting bit in all pmd that could map this
1280                  * hugepage, that will ensure no CPU can alter the
1281                  * mapcount on the head page. The mapcount is only
1282                  * accounted in the head page and it has to be
1283                  * transferred to all tail pages in the below code. So
1284                  * for this code to be safe, the split the mapcount
1285                  * can't change. But that doesn't mean userland can't
1286                  * keep changing and reading the page contents while
1287                  * we transfer the mapcount, so the pmd splitting
1288                  * status is achieved setting a reserved bit in the
1289                  * pmd, not by clearing the present bit.
1290                 */
1291                 page_tail->_mapcount = page->_mapcount;
1292
1293                 BUG_ON(page_tail->mapping);
1294                 page_tail->mapping = page->mapping;
1295
1296                 page_tail->index = page->index + i;
1297
1298                 BUG_ON(!PageAnon(page_tail));
1299                 BUG_ON(!PageUptodate(page_tail));
1300                 BUG_ON(!PageDirty(page_tail));
1301                 BUG_ON(!PageSwapBacked(page_tail));
1302
1303                 lru_add_page_tail(page, page_tail, lruvec);
1304         }
1305         atomic_sub(tail_count, &page->_count);
1306         BUG_ON(atomic_read(&page->_count) <= 0);
1307
1308         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1309         __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1310
1311         ClearPageCompound(page);
1312         compound_unlock(page);
1313         spin_unlock_irq(&zone->lru_lock);
1314
1315         for (i = 1; i < HPAGE_PMD_NR; i++) {
1316                 struct page *page_tail = page + i;
1317                 BUG_ON(page_count(page_tail) <= 0);
1318                 /*
1319                  * Tail pages may be freed if there wasn't any mapping
1320                  * like if add_to_swap() is running on a lru page that
1321                  * had its mapping zapped. And freeing these pages
1322                  * requires taking the lru_lock so we do the put_page
1323                  * of the tail pages after the split is complete.
1324                  */
1325                 put_page(page_tail);
1326         }
1327
1328         /*
1329          * Only the head page (now become a regular page) is required
1330          * to be pinned by the caller.
1331          */
1332         BUG_ON(page_count(page) <= 0);
1333 }
1334
1335 static int __split_huge_page_map(struct page *page,
1336                                  struct vm_area_struct *vma,
1337                                  unsigned long address)
1338 {
1339         struct mm_struct *mm = vma->vm_mm;
1340         pmd_t *pmd, _pmd;
1341         int ret = 0, i;
1342         pgtable_t pgtable;
1343         unsigned long haddr;
1344
1345         spin_lock(&mm->page_table_lock);
1346         pmd = page_check_address_pmd(page, mm, address,
1347                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1348         if (pmd) {
1349                 pgtable = pgtable_trans_huge_withdraw(mm);
1350                 pmd_populate(mm, &_pmd, pgtable);
1351
1352                 haddr = address;
1353                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1354                         pte_t *pte, entry;
1355                         BUG_ON(PageCompound(page+i));
1356                         entry = mk_pte(page + i, vma->vm_page_prot);
1357                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1358                         if (!pmd_write(*pmd))
1359                                 entry = pte_wrprotect(entry);
1360                         else
1361                                 BUG_ON(page_mapcount(page) != 1);
1362                         if (!pmd_young(*pmd))
1363                                 entry = pte_mkold(entry);
1364                         pte = pte_offset_map(&_pmd, haddr);
1365                         BUG_ON(!pte_none(*pte));
1366                         set_pte_at(mm, haddr, pte, entry);
1367                         pte_unmap(pte);
1368                 }
1369
1370                 smp_wmb(); /* make pte visible before pmd */
1371                 /*
1372                  * Up to this point the pmd is present and huge and
1373                  * userland has the whole access to the hugepage
1374                  * during the split (which happens in place). If we
1375                  * overwrite the pmd with the not-huge version
1376                  * pointing to the pte here (which of course we could
1377                  * if all CPUs were bug free), userland could trigger
1378                  * a small page size TLB miss on the small sized TLB
1379                  * while the hugepage TLB entry is still established
1380                  * in the huge TLB. Some CPU doesn't like that. See
1381                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1382                  * Erratum 383 on page 93. Intel should be safe but is
1383                  * also warns that it's only safe if the permission
1384                  * and cache attributes of the two entries loaded in
1385                  * the two TLB is identical (which should be the case
1386                  * here). But it is generally safer to never allow
1387                  * small and huge TLB entries for the same virtual
1388                  * address to be loaded simultaneously. So instead of
1389                  * doing "pmd_populate(); flush_tlb_range();" we first
1390                  * mark the current pmd notpresent (atomically because
1391                  * here the pmd_trans_huge and pmd_trans_splitting
1392                  * must remain set at all times on the pmd until the
1393                  * split is complete for this pmd), then we flush the
1394                  * SMP TLB and finally we write the non-huge version
1395                  * of the pmd entry with pmd_populate.
1396                  */
1397                 pmdp_invalidate(vma, address, pmd);
1398                 pmd_populate(mm, pmd, pgtable);
1399                 ret = 1;
1400         }
1401         spin_unlock(&mm->page_table_lock);
1402
1403         return ret;
1404 }
1405
1406 /* must be called with anon_vma->root->mutex hold */
1407 static void __split_huge_page(struct page *page,
1408                               struct anon_vma *anon_vma)
1409 {
1410         int mapcount, mapcount2;
1411         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1412         struct anon_vma_chain *avc;
1413
1414         BUG_ON(!PageHead(page));
1415         BUG_ON(PageTail(page));
1416
1417         mapcount = 0;
1418         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1419                 struct vm_area_struct *vma = avc->vma;
1420                 unsigned long addr = vma_address(page, vma);
1421                 BUG_ON(is_vma_temporary_stack(vma));
1422                 mapcount += __split_huge_page_splitting(page, vma, addr);
1423         }
1424         /*
1425          * It is critical that new vmas are added to the tail of the
1426          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1427          * and establishes a child pmd before
1428          * __split_huge_page_splitting() freezes the parent pmd (so if
1429          * we fail to prevent copy_huge_pmd() from running until the
1430          * whole __split_huge_page() is complete), we will still see
1431          * the newly established pmd of the child later during the
1432          * walk, to be able to set it as pmd_trans_splitting too.
1433          */
1434         if (mapcount != page_mapcount(page))
1435                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1436                        mapcount, page_mapcount(page));
1437         BUG_ON(mapcount != page_mapcount(page));
1438
1439         __split_huge_page_refcount(page);
1440
1441         mapcount2 = 0;
1442         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1443                 struct vm_area_struct *vma = avc->vma;
1444                 unsigned long addr = vma_address(page, vma);
1445                 BUG_ON(is_vma_temporary_stack(vma));
1446                 mapcount2 += __split_huge_page_map(page, vma, addr);
1447         }
1448         if (mapcount != mapcount2)
1449                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1450                        mapcount, mapcount2, page_mapcount(page));
1451         BUG_ON(mapcount != mapcount2);
1452 }
1453
1454 int split_huge_page(struct page *page)
1455 {
1456         struct anon_vma *anon_vma;
1457         int ret = 1;
1458
1459         BUG_ON(!PageAnon(page));
1460         anon_vma = page_lock_anon_vma(page);
1461         if (!anon_vma)
1462                 goto out;
1463         ret = 0;
1464         if (!PageCompound(page))
1465                 goto out_unlock;
1466
1467         BUG_ON(!PageSwapBacked(page));
1468         __split_huge_page(page, anon_vma);
1469         count_vm_event(THP_SPLIT);
1470
1471         BUG_ON(PageCompound(page));
1472 out_unlock:
1473         page_unlock_anon_vma(anon_vma);
1474 out:
1475         return ret;
1476 }
1477
1478 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1479
1480 int hugepage_madvise(struct vm_area_struct *vma,
1481                      unsigned long *vm_flags, int advice)
1482 {
1483         struct mm_struct *mm = vma->vm_mm;
1484
1485         switch (advice) {
1486         case MADV_HUGEPAGE:
1487                 /*
1488                  * Be somewhat over-protective like KSM for now!
1489                  */
1490                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1491                         return -EINVAL;
1492                 if (mm->def_flags & VM_NOHUGEPAGE)
1493                         return -EINVAL;
1494                 *vm_flags &= ~VM_NOHUGEPAGE;
1495                 *vm_flags |= VM_HUGEPAGE;
1496                 /*
1497                  * If the vma become good for khugepaged to scan,
1498                  * register it here without waiting a page fault that
1499                  * may not happen any time soon.
1500                  */
1501                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1502                         return -ENOMEM;
1503                 break;
1504         case MADV_NOHUGEPAGE:
1505                 /*
1506                  * Be somewhat over-protective like KSM for now!
1507                  */
1508                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1509                         return -EINVAL;
1510                 *vm_flags &= ~VM_HUGEPAGE;
1511                 *vm_flags |= VM_NOHUGEPAGE;
1512                 /*
1513                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1514                  * this vma even if we leave the mm registered in khugepaged if
1515                  * it got registered before VM_NOHUGEPAGE was set.
1516                  */
1517                 break;
1518         }
1519
1520         return 0;
1521 }
1522
1523 static int __init khugepaged_slab_init(void)
1524 {
1525         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1526                                           sizeof(struct mm_slot),
1527                                           __alignof__(struct mm_slot), 0, NULL);
1528         if (!mm_slot_cache)
1529                 return -ENOMEM;
1530
1531         return 0;
1532 }
1533
1534 static void __init khugepaged_slab_free(void)
1535 {
1536         kmem_cache_destroy(mm_slot_cache);
1537         mm_slot_cache = NULL;
1538 }
1539
1540 static inline struct mm_slot *alloc_mm_slot(void)
1541 {
1542         if (!mm_slot_cache)     /* initialization failed */
1543                 return NULL;
1544         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1545 }
1546
1547 static inline void free_mm_slot(struct mm_slot *mm_slot)
1548 {
1549         kmem_cache_free(mm_slot_cache, mm_slot);
1550 }
1551
1552 static int __init mm_slots_hash_init(void)
1553 {
1554         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1555                                 GFP_KERNEL);
1556         if (!mm_slots_hash)
1557                 return -ENOMEM;
1558         return 0;
1559 }
1560
1561 #if 0
1562 static void __init mm_slots_hash_free(void)
1563 {
1564         kfree(mm_slots_hash);
1565         mm_slots_hash = NULL;
1566 }
1567 #endif
1568
1569 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1570 {
1571         struct mm_slot *mm_slot;
1572         struct hlist_head *bucket;
1573         struct hlist_node *node;
1574
1575         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1576                                 % MM_SLOTS_HASH_HEADS];
1577         hlist_for_each_entry(mm_slot, node, bucket, hash) {
1578                 if (mm == mm_slot->mm)
1579                         return mm_slot;
1580         }
1581         return NULL;
1582 }
1583
1584 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1585                                     struct mm_slot *mm_slot)
1586 {
1587         struct hlist_head *bucket;
1588
1589         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1590                                 % MM_SLOTS_HASH_HEADS];
1591         mm_slot->mm = mm;
1592         hlist_add_head(&mm_slot->hash, bucket);
1593 }
1594
1595 static inline int khugepaged_test_exit(struct mm_struct *mm)
1596 {
1597         return atomic_read(&mm->mm_users) == 0;
1598 }
1599
1600 int __khugepaged_enter(struct mm_struct *mm)
1601 {
1602         struct mm_slot *mm_slot;
1603         int wakeup;
1604
1605         mm_slot = alloc_mm_slot();
1606         if (!mm_slot)
1607                 return -ENOMEM;
1608
1609         /* __khugepaged_exit() must not run from under us */
1610         VM_BUG_ON(khugepaged_test_exit(mm));
1611         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1612                 free_mm_slot(mm_slot);
1613                 return 0;
1614         }
1615
1616         spin_lock(&khugepaged_mm_lock);
1617         insert_to_mm_slots_hash(mm, mm_slot);
1618         /*
1619          * Insert just behind the scanning cursor, to let the area settle
1620          * down a little.
1621          */
1622         wakeup = list_empty(&khugepaged_scan.mm_head);
1623         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1624         spin_unlock(&khugepaged_mm_lock);
1625
1626         atomic_inc(&mm->mm_count);
1627         if (wakeup)
1628                 wake_up_interruptible(&khugepaged_wait);
1629
1630         return 0;
1631 }
1632
1633 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1634 {
1635         unsigned long hstart, hend;
1636         if (!vma->anon_vma)
1637                 /*
1638                  * Not yet faulted in so we will register later in the
1639                  * page fault if needed.
1640                  */
1641                 return 0;
1642         if (vma->vm_ops)
1643                 /* khugepaged not yet working on file or special mappings */
1644                 return 0;
1645         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
1646         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1647         hend = vma->vm_end & HPAGE_PMD_MASK;
1648         if (hstart < hend)
1649                 return khugepaged_enter(vma);
1650         return 0;
1651 }
1652
1653 void __khugepaged_exit(struct mm_struct *mm)
1654 {
1655         struct mm_slot *mm_slot;
1656         int free = 0;
1657
1658         spin_lock(&khugepaged_mm_lock);
1659         mm_slot = get_mm_slot(mm);
1660         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1661                 hlist_del(&mm_slot->hash);
1662                 list_del(&mm_slot->mm_node);
1663                 free = 1;
1664         }
1665         spin_unlock(&khugepaged_mm_lock);
1666
1667         if (free) {
1668                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1669                 free_mm_slot(mm_slot);
1670                 mmdrop(mm);
1671         } else if (mm_slot) {
1672                 /*
1673                  * This is required to serialize against
1674                  * khugepaged_test_exit() (which is guaranteed to run
1675                  * under mmap sem read mode). Stop here (after we
1676                  * return all pagetables will be destroyed) until
1677                  * khugepaged has finished working on the pagetables
1678                  * under the mmap_sem.
1679                  */
1680                 down_write(&mm->mmap_sem);
1681                 up_write(&mm->mmap_sem);
1682         }
1683 }
1684
1685 static void release_pte_page(struct page *page)
1686 {
1687         /* 0 stands for page_is_file_cache(page) == false */
1688         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1689         unlock_page(page);
1690         putback_lru_page(page);
1691 }
1692
1693 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1694 {
1695         while (--_pte >= pte) {
1696                 pte_t pteval = *_pte;
1697                 if (!pte_none(pteval))
1698                         release_pte_page(pte_page(pteval));
1699         }
1700 }
1701
1702 static void release_all_pte_pages(pte_t *pte)
1703 {
1704         release_pte_pages(pte, pte + HPAGE_PMD_NR);
1705 }
1706
1707 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1708                                         unsigned long address,
1709                                         pte_t *pte)
1710 {
1711         struct page *page;
1712         pte_t *_pte;
1713         int referenced = 0, isolated = 0, none = 0;
1714         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1715              _pte++, address += PAGE_SIZE) {
1716                 pte_t pteval = *_pte;
1717                 if (pte_none(pteval)) {
1718                         if (++none <= khugepaged_max_ptes_none)
1719                                 continue;
1720                         else {
1721                                 release_pte_pages(pte, _pte);
1722                                 goto out;
1723                         }
1724                 }
1725                 if (!pte_present(pteval) || !pte_write(pteval)) {
1726                         release_pte_pages(pte, _pte);
1727                         goto out;
1728                 }
1729                 page = vm_normal_page(vma, address, pteval);
1730                 if (unlikely(!page)) {
1731                         release_pte_pages(pte, _pte);
1732                         goto out;
1733                 }
1734                 VM_BUG_ON(PageCompound(page));
1735                 BUG_ON(!PageAnon(page));
1736                 VM_BUG_ON(!PageSwapBacked(page));
1737
1738                 /* cannot use mapcount: can't collapse if there's a gup pin */
1739                 if (page_count(page) != 1) {
1740                         release_pte_pages(pte, _pte);
1741                         goto out;
1742                 }
1743                 /*
1744                  * We can do it before isolate_lru_page because the
1745                  * page can't be freed from under us. NOTE: PG_lock
1746                  * is needed to serialize against split_huge_page
1747                  * when invoked from the VM.
1748                  */
1749                 if (!trylock_page(page)) {
1750                         release_pte_pages(pte, _pte);
1751                         goto out;
1752                 }
1753                 /*
1754                  * Isolate the page to avoid collapsing an hugepage
1755                  * currently in use by the VM.
1756                  */
1757                 if (isolate_lru_page(page)) {
1758                         unlock_page(page);
1759                         release_pte_pages(pte, _pte);
1760                         goto out;
1761                 }
1762                 /* 0 stands for page_is_file_cache(page) == false */
1763                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1764                 VM_BUG_ON(!PageLocked(page));
1765                 VM_BUG_ON(PageLRU(page));
1766
1767                 /* If there is no mapped pte young don't collapse the page */
1768                 if (pte_young(pteval) || PageReferenced(page) ||
1769                     mmu_notifier_test_young(vma->vm_mm, address))
1770                         referenced = 1;
1771         }
1772         if (unlikely(!referenced))
1773                 release_all_pte_pages(pte);
1774         else
1775                 isolated = 1;
1776 out:
1777         return isolated;
1778 }
1779
1780 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1781                                       struct vm_area_struct *vma,
1782                                       unsigned long address,
1783                                       spinlock_t *ptl)
1784 {
1785         pte_t *_pte;
1786         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1787                 pte_t pteval = *_pte;
1788                 struct page *src_page;
1789
1790                 if (pte_none(pteval)) {
1791                         clear_user_highpage(page, address);
1792                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1793                 } else {
1794                         src_page = pte_page(pteval);
1795                         copy_user_highpage(page, src_page, address, vma);
1796                         VM_BUG_ON(page_mapcount(src_page) != 1);
1797                         release_pte_page(src_page);
1798                         /*
1799                          * ptl mostly unnecessary, but preempt has to
1800                          * be disabled to update the per-cpu stats
1801                          * inside page_remove_rmap().
1802                          */
1803                         spin_lock(ptl);
1804                         /*
1805                          * paravirt calls inside pte_clear here are
1806                          * superfluous.
1807                          */
1808                         pte_clear(vma->vm_mm, address, _pte);
1809                         page_remove_rmap(src_page);
1810                         spin_unlock(ptl);
1811                         free_page_and_swap_cache(src_page);
1812                 }
1813
1814                 address += PAGE_SIZE;
1815                 page++;
1816         }
1817 }
1818
1819 static void khugepaged_alloc_sleep(void)
1820 {
1821         wait_event_freezable_timeout(khugepaged_wait, false,
1822                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
1823 }
1824
1825 #ifdef CONFIG_NUMA
1826 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
1827 {
1828         if (IS_ERR(*hpage)) {
1829                 if (!*wait)
1830                         return false;
1831
1832                 *wait = false;
1833                 *hpage = NULL;
1834                 khugepaged_alloc_sleep();
1835         } else if (*hpage) {
1836                 put_page(*hpage);
1837                 *hpage = NULL;
1838         }
1839
1840         return true;
1841 }
1842
1843 static struct page
1844 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
1845                        struct vm_area_struct *vma, unsigned long address,
1846                        int node)
1847 {
1848         VM_BUG_ON(*hpage);
1849         /*
1850          * Allocate the page while the vma is still valid and under
1851          * the mmap_sem read mode so there is no memory allocation
1852          * later when we take the mmap_sem in write mode. This is more
1853          * friendly behavior (OTOH it may actually hide bugs) to
1854          * filesystems in userland with daemons allocating memory in
1855          * the userland I/O paths.  Allocating memory with the
1856          * mmap_sem in read mode is good idea also to allow greater
1857          * scalability.
1858          */
1859         *hpage  = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
1860                                       node, __GFP_OTHER_NODE);
1861
1862         /*
1863          * After allocating the hugepage, release the mmap_sem read lock in
1864          * preparation for taking it in write mode.
1865          */
1866         up_read(&mm->mmap_sem);
1867         if (unlikely(!*hpage)) {
1868                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1869                 *hpage = ERR_PTR(-ENOMEM);
1870                 return NULL;
1871         }
1872
1873         count_vm_event(THP_COLLAPSE_ALLOC);
1874         return *hpage;
1875 }
1876 #else
1877 static struct page *khugepaged_alloc_hugepage(bool *wait)
1878 {
1879         struct page *hpage;
1880
1881         do {
1882                 hpage = alloc_hugepage(khugepaged_defrag());
1883                 if (!hpage) {
1884                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1885                         if (!*wait)
1886                                 return NULL;
1887
1888                         *wait = false;
1889                         khugepaged_alloc_sleep();
1890                 } else
1891                         count_vm_event(THP_COLLAPSE_ALLOC);
1892         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
1893
1894         return hpage;
1895 }
1896
1897 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
1898 {
1899         if (!*hpage)
1900                 *hpage = khugepaged_alloc_hugepage(wait);
1901
1902         if (unlikely(!*hpage))
1903                 return false;
1904
1905         return true;
1906 }
1907
1908 static struct page
1909 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
1910                        struct vm_area_struct *vma, unsigned long address,
1911                        int node)
1912 {
1913         up_read(&mm->mmap_sem);
1914         VM_BUG_ON(!*hpage);
1915         return  *hpage;
1916 }
1917 #endif
1918
1919 static void collapse_huge_page(struct mm_struct *mm,
1920                                    unsigned long address,
1921                                    struct page **hpage,
1922                                    struct vm_area_struct *vma,
1923                                    int node)
1924 {
1925         pgd_t *pgd;
1926         pud_t *pud;
1927         pmd_t *pmd, _pmd;
1928         pte_t *pte;
1929         pgtable_t pgtable;
1930         struct page *new_page;
1931         spinlock_t *ptl;
1932         int isolated;
1933         unsigned long hstart, hend;
1934         unsigned long mmun_start;       /* For mmu_notifiers */
1935         unsigned long mmun_end;         /* For mmu_notifiers */
1936
1937         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1938
1939         /* release the mmap_sem read lock. */
1940         new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
1941         if (!new_page)
1942                 return;
1943
1944         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
1945                 return;
1946
1947         /*
1948          * Prevent all access to pagetables with the exception of
1949          * gup_fast later hanlded by the ptep_clear_flush and the VM
1950          * handled by the anon_vma lock + PG_lock.
1951          */
1952         down_write(&mm->mmap_sem);
1953         if (unlikely(khugepaged_test_exit(mm)))
1954                 goto out;
1955
1956         vma = find_vma(mm, address);
1957         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1958         hend = vma->vm_end & HPAGE_PMD_MASK;
1959         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
1960                 goto out;
1961
1962         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
1963             (vma->vm_flags & VM_NOHUGEPAGE))
1964                 goto out;
1965
1966         if (!vma->anon_vma || vma->vm_ops)
1967                 goto out;
1968         if (is_vma_temporary_stack(vma))
1969                 goto out;
1970         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
1971
1972         pgd = pgd_offset(mm, address);
1973         if (!pgd_present(*pgd))
1974                 goto out;
1975
1976         pud = pud_offset(pgd, address);
1977         if (!pud_present(*pud))
1978                 goto out;
1979
1980         pmd = pmd_offset(pud, address);
1981         /* pmd can't go away or become huge under us */
1982         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1983                 goto out;
1984
1985         anon_vma_lock(vma->anon_vma);
1986
1987         pte = pte_offset_map(pmd, address);
1988         ptl = pte_lockptr(mm, pmd);
1989
1990         mmun_start = address;
1991         mmun_end   = address + HPAGE_PMD_SIZE;
1992         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1993         spin_lock(&mm->page_table_lock); /* probably unnecessary */
1994         /*
1995          * After this gup_fast can't run anymore. This also removes
1996          * any huge TLB entry from the CPU so we won't allow
1997          * huge and small TLB entries for the same virtual address
1998          * to avoid the risk of CPU bugs in that area.
1999          */
2000         _pmd = pmdp_clear_flush(vma, address, pmd);
2001         spin_unlock(&mm->page_table_lock);
2002         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2003
2004         spin_lock(ptl);
2005         isolated = __collapse_huge_page_isolate(vma, address, pte);
2006         spin_unlock(ptl);
2007
2008         if (unlikely(!isolated)) {
2009                 pte_unmap(pte);
2010                 spin_lock(&mm->page_table_lock);
2011                 BUG_ON(!pmd_none(*pmd));
2012                 set_pmd_at(mm, address, pmd, _pmd);
2013                 spin_unlock(&mm->page_table_lock);
2014                 anon_vma_unlock(vma->anon_vma);
2015                 goto out;
2016         }
2017
2018         /*
2019          * All pages are isolated and locked so anon_vma rmap
2020          * can't run anymore.
2021          */
2022         anon_vma_unlock(vma->anon_vma);
2023
2024         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
2025         pte_unmap(pte);
2026         __SetPageUptodate(new_page);
2027         pgtable = pmd_pgtable(_pmd);
2028
2029         _pmd = mk_pmd(new_page, vma->vm_page_prot);
2030         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2031         _pmd = pmd_mkhuge(_pmd);
2032
2033         /*
2034          * spin_lock() below is not the equivalent of smp_wmb(), so
2035          * this is needed to avoid the copy_huge_page writes to become
2036          * visible after the set_pmd_at() write.
2037          */
2038         smp_wmb();
2039
2040         spin_lock(&mm->page_table_lock);
2041         BUG_ON(!pmd_none(*pmd));
2042         page_add_new_anon_rmap(new_page, vma, address);
2043         set_pmd_at(mm, address, pmd, _pmd);
2044         update_mmu_cache(vma, address, pmd);
2045         pgtable_trans_huge_deposit(mm, pgtable);
2046         spin_unlock(&mm->page_table_lock);
2047
2048         *hpage = NULL;
2049
2050         khugepaged_pages_collapsed++;
2051 out_up_write:
2052         up_write(&mm->mmap_sem);
2053         return;
2054
2055 out:
2056         mem_cgroup_uncharge_page(new_page);
2057         goto out_up_write;
2058 }
2059
2060 static int khugepaged_scan_pmd(struct mm_struct *mm,
2061                                struct vm_area_struct *vma,
2062                                unsigned long address,
2063                                struct page **hpage)
2064 {
2065         pgd_t *pgd;
2066         pud_t *pud;
2067         pmd_t *pmd;
2068         pte_t *pte, *_pte;
2069         int ret = 0, referenced = 0, none = 0;
2070         struct page *page;
2071         unsigned long _address;
2072         spinlock_t *ptl;
2073         int node = -1;
2074
2075         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2076
2077         pgd = pgd_offset(mm, address);
2078         if (!pgd_present(*pgd))
2079                 goto out;
2080
2081         pud = pud_offset(pgd, address);
2082         if (!pud_present(*pud))
2083                 goto out;
2084
2085         pmd = pmd_offset(pud, address);
2086         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2087                 goto out;
2088
2089         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2090         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2091              _pte++, _address += PAGE_SIZE) {
2092                 pte_t pteval = *_pte;
2093                 if (pte_none(pteval)) {
2094                         if (++none <= khugepaged_max_ptes_none)
2095                                 continue;
2096                         else
2097                                 goto out_unmap;
2098                 }
2099                 if (!pte_present(pteval) || !pte_write(pteval))
2100                         goto out_unmap;
2101                 page = vm_normal_page(vma, _address, pteval);
2102                 if (unlikely(!page))
2103                         goto out_unmap;
2104                 /*
2105                  * Chose the node of the first page. This could
2106                  * be more sophisticated and look at more pages,
2107                  * but isn't for now.
2108                  */
2109                 if (node == -1)
2110                         node = page_to_nid(page);
2111                 VM_BUG_ON(PageCompound(page));
2112                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2113                         goto out_unmap;
2114                 /* cannot use mapcount: can't collapse if there's a gup pin */
2115                 if (page_count(page) != 1)
2116                         goto out_unmap;
2117                 if (pte_young(pteval) || PageReferenced(page) ||
2118                     mmu_notifier_test_young(vma->vm_mm, address))
2119                         referenced = 1;
2120         }
2121         if (referenced)
2122                 ret = 1;
2123 out_unmap:
2124         pte_unmap_unlock(pte, ptl);
2125         if (ret)
2126                 /* collapse_huge_page will return with the mmap_sem released */
2127                 collapse_huge_page(mm, address, hpage, vma, node);
2128 out:
2129         return ret;
2130 }
2131
2132 static void collect_mm_slot(struct mm_slot *mm_slot)
2133 {
2134         struct mm_struct *mm = mm_slot->mm;
2135
2136         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2137
2138         if (khugepaged_test_exit(mm)) {
2139                 /* free mm_slot */
2140                 hlist_del(&mm_slot->hash);
2141                 list_del(&mm_slot->mm_node);
2142
2143                 /*
2144                  * Not strictly needed because the mm exited already.
2145                  *
2146                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2147                  */
2148
2149                 /* khugepaged_mm_lock actually not necessary for the below */
2150                 free_mm_slot(mm_slot);
2151                 mmdrop(mm);
2152         }
2153 }
2154
2155 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2156                                             struct page **hpage)
2157         __releases(&khugepaged_mm_lock)
2158         __acquires(&khugepaged_mm_lock)
2159 {
2160         struct mm_slot *mm_slot;
2161         struct mm_struct *mm;
2162         struct vm_area_struct *vma;
2163         int progress = 0;
2164
2165         VM_BUG_ON(!pages);
2166         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2167
2168         if (khugepaged_scan.mm_slot)
2169                 mm_slot = khugepaged_scan.mm_slot;
2170         else {
2171                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2172                                      struct mm_slot, mm_node);
2173                 khugepaged_scan.address = 0;
2174                 khugepaged_scan.mm_slot = mm_slot;
2175         }
2176         spin_unlock(&khugepaged_mm_lock);
2177
2178         mm = mm_slot->mm;
2179         down_read(&mm->mmap_sem);
2180         if (unlikely(khugepaged_test_exit(mm)))
2181                 vma = NULL;
2182         else
2183                 vma = find_vma(mm, khugepaged_scan.address);
2184
2185         progress++;
2186         for (; vma; vma = vma->vm_next) {
2187                 unsigned long hstart, hend;
2188
2189                 cond_resched();
2190                 if (unlikely(khugepaged_test_exit(mm))) {
2191                         progress++;
2192                         break;
2193                 }
2194
2195                 if ((!(vma->vm_flags & VM_HUGEPAGE) &&
2196                      !khugepaged_always()) ||
2197                     (vma->vm_flags & VM_NOHUGEPAGE)) {
2198                 skip:
2199                         progress++;
2200                         continue;
2201                 }
2202                 if (!vma->anon_vma || vma->vm_ops)
2203                         goto skip;
2204                 if (is_vma_temporary_stack(vma))
2205                         goto skip;
2206                 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2207
2208                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2209                 hend = vma->vm_end & HPAGE_PMD_MASK;
2210                 if (hstart >= hend)
2211                         goto skip;
2212                 if (khugepaged_scan.address > hend)
2213                         goto skip;
2214                 if (khugepaged_scan.address < hstart)
2215                         khugepaged_scan.address = hstart;
2216                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2217
2218                 while (khugepaged_scan.address < hend) {
2219                         int ret;
2220                         cond_resched();
2221                         if (unlikely(khugepaged_test_exit(mm)))
2222                                 goto breakouterloop;
2223
2224                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2225                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2226                                   hend);
2227                         ret = khugepaged_scan_pmd(mm, vma,
2228                                                   khugepaged_scan.address,
2229                                                   hpage);
2230                         /* move to next address */
2231                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2232                         progress += HPAGE_PMD_NR;
2233                         if (ret)
2234                                 /* we released mmap_sem so break loop */
2235                                 goto breakouterloop_mmap_sem;
2236                         if (progress >= pages)
2237                                 goto breakouterloop;
2238                 }
2239         }
2240 breakouterloop:
2241         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2242 breakouterloop_mmap_sem:
2243
2244         spin_lock(&khugepaged_mm_lock);
2245         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2246         /*
2247          * Release the current mm_slot if this mm is about to die, or
2248          * if we scanned all vmas of this mm.
2249          */
2250         if (khugepaged_test_exit(mm) || !vma) {
2251                 /*
2252                  * Make sure that if mm_users is reaching zero while
2253                  * khugepaged runs here, khugepaged_exit will find
2254                  * mm_slot not pointing to the exiting mm.
2255                  */
2256                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2257                         khugepaged_scan.mm_slot = list_entry(
2258                                 mm_slot->mm_node.next,
2259                                 struct mm_slot, mm_node);
2260                         khugepaged_scan.address = 0;
2261                 } else {
2262                         khugepaged_scan.mm_slot = NULL;
2263                         khugepaged_full_scans++;
2264                 }
2265
2266                 collect_mm_slot(mm_slot);
2267         }
2268
2269         return progress;
2270 }
2271
2272 static int khugepaged_has_work(void)
2273 {
2274         return !list_empty(&khugepaged_scan.mm_head) &&
2275                 khugepaged_enabled();
2276 }
2277
2278 static int khugepaged_wait_event(void)
2279 {
2280         return !list_empty(&khugepaged_scan.mm_head) ||
2281                 kthread_should_stop();
2282 }
2283
2284 static void khugepaged_do_scan(void)
2285 {
2286         struct page *hpage = NULL;
2287         unsigned int progress = 0, pass_through_head = 0;
2288         unsigned int pages = khugepaged_pages_to_scan;
2289         bool wait = true;
2290
2291         barrier(); /* write khugepaged_pages_to_scan to local stack */
2292
2293         while (progress < pages) {
2294                 if (!khugepaged_prealloc_page(&hpage, &wait))
2295                         break;
2296
2297                 cond_resched();
2298
2299                 if (unlikely(kthread_should_stop() || freezing(current)))
2300                         break;
2301
2302                 spin_lock(&khugepaged_mm_lock);
2303                 if (!khugepaged_scan.mm_slot)
2304                         pass_through_head++;
2305                 if (khugepaged_has_work() &&
2306                     pass_through_head < 2)
2307                         progress += khugepaged_scan_mm_slot(pages - progress,
2308                                                             &hpage);
2309                 else
2310                         progress = pages;
2311                 spin_unlock(&khugepaged_mm_lock);
2312         }
2313
2314         if (!IS_ERR_OR_NULL(hpage))
2315                 put_page(hpage);
2316 }
2317
2318 static void khugepaged_wait_work(void)
2319 {
2320         try_to_freeze();
2321
2322         if (khugepaged_has_work()) {
2323                 if (!khugepaged_scan_sleep_millisecs)
2324                         return;
2325
2326                 wait_event_freezable_timeout(khugepaged_wait,
2327                                              kthread_should_stop(),
2328                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2329                 return;
2330         }
2331
2332         if (khugepaged_enabled())
2333                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2334 }
2335
2336 static int khugepaged(void *none)
2337 {
2338         struct mm_slot *mm_slot;
2339
2340         set_freezable();
2341         set_user_nice(current, 19);
2342
2343         while (!kthread_should_stop()) {
2344                 khugepaged_do_scan();
2345                 khugepaged_wait_work();
2346         }
2347
2348         spin_lock(&khugepaged_mm_lock);
2349         mm_slot = khugepaged_scan.mm_slot;
2350         khugepaged_scan.mm_slot = NULL;
2351         if (mm_slot)
2352                 collect_mm_slot(mm_slot);
2353         spin_unlock(&khugepaged_mm_lock);
2354         return 0;
2355 }
2356
2357 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2358 {
2359         struct page *page;
2360
2361         spin_lock(&mm->page_table_lock);
2362         if (unlikely(!pmd_trans_huge(*pmd))) {
2363                 spin_unlock(&mm->page_table_lock);
2364                 return;
2365         }
2366         page = pmd_page(*pmd);
2367         VM_BUG_ON(!page_count(page));
2368         get_page(page);
2369         spin_unlock(&mm->page_table_lock);
2370
2371         split_huge_page(page);
2372
2373         put_page(page);
2374         BUG_ON(pmd_trans_huge(*pmd));
2375 }
2376
2377 static void split_huge_page_address(struct mm_struct *mm,
2378                                     unsigned long address)
2379 {
2380         pgd_t *pgd;
2381         pud_t *pud;
2382         pmd_t *pmd;
2383
2384         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2385
2386         pgd = pgd_offset(mm, address);
2387         if (!pgd_present(*pgd))
2388                 return;
2389
2390         pud = pud_offset(pgd, address);
2391         if (!pud_present(*pud))
2392                 return;
2393
2394         pmd = pmd_offset(pud, address);
2395         if (!pmd_present(*pmd))
2396                 return;
2397         /*
2398          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2399          * materialize from under us.
2400          */
2401         split_huge_page_pmd(mm, pmd);
2402 }
2403
2404 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2405                              unsigned long start,
2406                              unsigned long end,
2407                              long adjust_next)
2408 {
2409         /*
2410          * If the new start address isn't hpage aligned and it could
2411          * previously contain an hugepage: check if we need to split
2412          * an huge pmd.
2413          */
2414         if (start & ~HPAGE_PMD_MASK &&
2415             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2416             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2417                 split_huge_page_address(vma->vm_mm, start);
2418
2419         /*
2420          * If the new end address isn't hpage aligned and it could
2421          * previously contain an hugepage: check if we need to split
2422          * an huge pmd.
2423          */
2424         if (end & ~HPAGE_PMD_MASK &&
2425             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2426             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2427                 split_huge_page_address(vma->vm_mm, end);
2428
2429         /*
2430          * If we're also updating the vma->vm_next->vm_start, if the new
2431          * vm_next->vm_start isn't page aligned and it could previously
2432          * contain an hugepage: check if we need to split an huge pmd.
2433          */
2434         if (adjust_next > 0) {
2435                 struct vm_area_struct *next = vma->vm_next;
2436                 unsigned long nstart = next->vm_start;
2437                 nstart += adjust_next << PAGE_SHIFT;
2438                 if (nstart & ~HPAGE_PMD_MASK &&
2439                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2440                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2441                         split_huge_page_address(next->vm_mm, nstart);
2442         }
2443 }