mm/khugepaged.c

   1 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   2
   3 #include <linux/mm.h>
   4 #include <linux/sched.h>
   5 #include <linux/mmu_notifier.h>
   6 #include <linux/rmap.h>
   7 #include <linux/swap.h>
   8 #include <linux/mm_inline.h>
   9 #include <linux/kthread.h>
  10 #include <linux/khugepaged.h>
  11 #include <linux/freezer.h>
  12 #include <linux/mman.h>
  13 #include <linux/hashtable.h>
  14 #include <linux/userfaultfd_k.h>
  15 #include <linux/page_idle.h>
  16 #include <linux/swapops.h>
  17
  18 #include <asm/tlb.h>
  19 #include <asm/pgalloc.h>
  20 #include "internal.h"
  21
  22 enum scan_result {
  23         SCAN_FAIL,
  24         SCAN_SUCCEED,
  25         SCAN_PMD_NULL,
  26         SCAN_EXCEED_NONE_PTE,
  27         SCAN_PTE_NON_PRESENT,
  28         SCAN_PAGE_RO,
  29         SCAN_NO_REFERENCED_PAGE,
  30         SCAN_PAGE_NULL,
  31         SCAN_SCAN_ABORT,
  32         SCAN_PAGE_COUNT,
  33         SCAN_PAGE_LRU,
  34         SCAN_PAGE_LOCK,
  35         SCAN_PAGE_ANON,
  36         SCAN_PAGE_COMPOUND,
  37         SCAN_ANY_PROCESS,
  38         SCAN_VMA_NULL,
  39         SCAN_VMA_CHECK,
  40         SCAN_ADDRESS_RANGE,
  41         SCAN_SWAP_CACHE_PAGE,
  42         SCAN_DEL_PAGE_LRU,
  43         SCAN_ALLOC_HUGE_PAGE_FAIL,
  44         SCAN_CGROUP_CHARGE_FAIL,
  45         SCAN_EXCEED_SWAP_PTE
  46 };
  47
  48 #define CREATE_TRACE_POINTS
  49 #include <trace/events/huge_memory.h>
  50
  51 /* default scan 8*512 pte (or vmas) every 30 second */
  52 static unsigned int khugepaged_pages_to_scan __read_mostly;
  53 static unsigned int khugepaged_pages_collapsed;
  54 static unsigned int khugepaged_full_scans;
  55 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  56 /* during fragmentation poll the hugepage allocator once every minute */
  57 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  58 static unsigned long khugepaged_sleep_expire;
  59 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  60 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  61 /*
  62  * default collapse hugepages if there is at least one pte mapped like
  63  * it would have happened if the vma was large enough during page
  64  * fault.
  65  */
  66 static unsigned int khugepaged_max_ptes_none __read_mostly;
  67 static unsigned int khugepaged_max_ptes_swap __read_mostly;
  68
  69 #define MM_SLOTS_HASH_BITS 10
  70 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  71
  72 static struct kmem_cache *mm_slot_cache __read_mostly;
  73
  74 /**
  75  * struct mm_slot - hash lookup from mm to mm_slot
  76  * @hash: hash collision list
  77  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  78  * @mm: the mm that this information is valid for
  79  */
  80 struct mm_slot {
  81         struct hlist_node hash;
  82         struct list_head mm_node;
  83         struct mm_struct *mm;
  84 };
  85
  86 /**
  87  * struct khugepaged_scan - cursor for scanning
  88  * @mm_head: the head of the mm list to scan
  89  * @mm_slot: the current mm_slot we are scanning
  90  * @address: the next address inside that to be scanned
  91  *
  92  * There is only the one khugepaged_scan instance of this cursor structure.
  93  */
  94 struct khugepaged_scan {
  95         struct list_head mm_head;
  96         struct mm_slot *mm_slot;
  97         unsigned long address;
  98 };
  99
 100 static struct khugepaged_scan khugepaged_scan = {
 101         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 102 };
 103
 104 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 105                                          struct kobj_attribute *attr,
 106                                          char *buf)
 107 {
 108         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 109 }
 110
 111 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 112                                           struct kobj_attribute *attr,
 113                                           const char *buf, size_t count)
 114 {
 115         unsigned long msecs;
 116         int err;
 117
 118         err = kstrtoul(buf, 10, &msecs);
 119         if (err || msecs > UINT_MAX)
 120                 return -EINVAL;
 121
 122         khugepaged_scan_sleep_millisecs = msecs;
 123         khugepaged_sleep_expire = 0;
 124         wake_up_interruptible(&khugepaged_wait);
 125
 126         return count;
 127 }
 128 static struct kobj_attribute scan_sleep_millisecs_attr =
 129         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 130                scan_sleep_millisecs_store);
 131
 132 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 133                                           struct kobj_attribute *attr,
 134                                           char *buf)
 135 {
 136         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 137 }
 138
 139 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 140                                            struct kobj_attribute *attr,
 141                                            const char *buf, size_t count)
 142 {
 143         unsigned long msecs;
 144         int err;
 145
 146         err = kstrtoul(buf, 10, &msecs);
 147         if (err || msecs > UINT_MAX)
 148                 return -EINVAL;
 149
 150         khugepaged_alloc_sleep_millisecs = msecs;
 151         khugepaged_sleep_expire = 0;
 152         wake_up_interruptible(&khugepaged_wait);
 153
 154         return count;
 155 }
 156 static struct kobj_attribute alloc_sleep_millisecs_attr =
 157         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 158                alloc_sleep_millisecs_store);
 159
 160 static ssize_t pages_to_scan_show(struct kobject *kobj,
 161                                   struct kobj_attribute *attr,
 162                                   char *buf)
 163 {
 164         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 165 }
 166 static ssize_t pages_to_scan_store(struct kobject *kobj,
 167                                    struct kobj_attribute *attr,
 168                                    const char *buf, size_t count)
 169 {
 170         int err;
 171         unsigned long pages;
 172
 173         err = kstrtoul(buf, 10, &pages);
 174         if (err || !pages || pages > UINT_MAX)
 175                 return -EINVAL;
 176
 177         khugepaged_pages_to_scan = pages;
 178
 179         return count;
 180 }
 181 static struct kobj_attribute pages_to_scan_attr =
 182         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 183                pages_to_scan_store);
 184
 185 static ssize_t pages_collapsed_show(struct kobject *kobj,
 186                                     struct kobj_attribute *attr,
 187                                     char *buf)
 188 {
 189         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 190 }
 191 static struct kobj_attribute pages_collapsed_attr =
 192         __ATTR_RO(pages_collapsed);
 193
 194 static ssize_t full_scans_show(struct kobject *kobj,
 195                                struct kobj_attribute *attr,
 196                                char *buf)
 197 {
 198         return sprintf(buf, "%u\n", khugepaged_full_scans);
 199 }
 200 static struct kobj_attribute full_scans_attr =
 201         __ATTR_RO(full_scans);
 202
 203 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 204                                       struct kobj_attribute *attr, char *buf)
 205 {
 206         return single_hugepage_flag_show(kobj, attr, buf,
 207                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 208 }
 209 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 210                                        struct kobj_attribute *attr,
 211                                        const char *buf, size_t count)
 212 {
 213         return single_hugepage_flag_store(kobj, attr, buf, count,
 214                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 215 }
 216 static struct kobj_attribute khugepaged_defrag_attr =
 217         __ATTR(defrag, 0644, khugepaged_defrag_show,
 218                khugepaged_defrag_store);
 219
 220 /*
 221  * max_ptes_none controls if khugepaged should collapse hugepages over
 222  * any unmapped ptes in turn potentially increasing the memory
 223  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 224  * reduce the available free memory in the system as it
 225  * runs. Increasing max_ptes_none will instead potentially reduce the
 226  * free memory in the system during the khugepaged scan.
 227  */
 228 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 229                                              struct kobj_attribute *attr,
 230                                              char *buf)
 231 {
 232         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 233 }
 234 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 235                                               struct kobj_attribute *attr,
 236                                               const char *buf, size_t count)
 237 {
 238         int err;
 239         unsigned long max_ptes_none;
 240
 241         err = kstrtoul(buf, 10, &max_ptes_none);
 242         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 243                 return -EINVAL;
 244
 245         khugepaged_max_ptes_none = max_ptes_none;
 246
 247         return count;
 248 }
 249 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 250         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 251                khugepaged_max_ptes_none_store);
 252
 253 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
 254                                              struct kobj_attribute *attr,
 255                                              char *buf)
 256 {
 257         return sprintf(buf, "%u\n", khugepaged_max_ptes_swap);
 258 }
 259
 260 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
 261                                               struct kobj_attribute *attr,
 262                                               const char *buf, size_t count)
 263 {
 264         int err;
 265         unsigned long max_ptes_swap;
 266
 267         err  = kstrtoul(buf, 10, &max_ptes_swap);
 268         if (err || max_ptes_swap > HPAGE_PMD_NR-1)
 269                 return -EINVAL;
 270
 271         khugepaged_max_ptes_swap = max_ptes_swap;
 272
 273         return count;
 274 }
 275
 276 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
 277         __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
 278                khugepaged_max_ptes_swap_store);
 279
 280 static struct attribute *khugepaged_attr[] = {
 281         &khugepaged_defrag_attr.attr,
 282         &khugepaged_max_ptes_none_attr.attr,
 283         &pages_to_scan_attr.attr,
 284         &pages_collapsed_attr.attr,
 285         &full_scans_attr.attr,
 286         &scan_sleep_millisecs_attr.attr,
 287         &alloc_sleep_millisecs_attr.attr,
 288         &khugepaged_max_ptes_swap_attr.attr,
 289         NULL,
 290 };
 291
 292 struct attribute_group khugepaged_attr_group = {
 293         .attrs = khugepaged_attr,
 294         .name = "khugepaged",
 295 };
 296
 297 #define VM_NO_KHUGEPAGED (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
 298
 299 int hugepage_madvise(struct vm_area_struct *vma,
 300                      unsigned long *vm_flags, int advice)
 301 {
 302         switch (advice) {
 303         case MADV_HUGEPAGE:
 304 #ifdef CONFIG_S390
 305                 /*
 306                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
 307                  * can't handle this properly after s390_enable_sie, so we simply
 308                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
 309                  */
 310                 if (mm_has_pgste(vma->vm_mm))
 311                         return 0;
 312 #endif
 313                 *vm_flags &= ~VM_NOHUGEPAGE;
 314                 *vm_flags |= VM_HUGEPAGE;
 315                 /*
 316                  * If the vma become good for khugepaged to scan,
 317                  * register it here without waiting a page fault that
 318                  * may not happen any time soon.
 319                  */
 320                 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
 321                                 khugepaged_enter_vma_merge(vma, *vm_flags))
 322                         return -ENOMEM;
 323                 break;
 324         case MADV_NOHUGEPAGE:
 325                 *vm_flags &= ~VM_HUGEPAGE;
 326                 *vm_flags |= VM_NOHUGEPAGE;
 327                 /*
 328                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
 329                  * this vma even if we leave the mm registered in khugepaged if
 330                  * it got registered before VM_NOHUGEPAGE was set.
 331                  */
 332                 break;
 333         }
 334
 335         return 0;
 336 }
 337
 338 int __init khugepaged_init(void)
 339 {
 340         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
 341                                           sizeof(struct mm_slot),
 342                                           __alignof__(struct mm_slot), 0, NULL);
 343         if (!mm_slot_cache)
 344                 return -ENOMEM;
 345
 346         khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
 347         khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
 348         khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
 349
 350         return 0;
 351 }
 352
 353 void __init khugepaged_destroy(void)
 354 {
 355         kmem_cache_destroy(mm_slot_cache);
 356 }
 357
 358 static inline struct mm_slot *alloc_mm_slot(void)
 359 {
 360         if (!mm_slot_cache)     /* initialization failed */
 361                 return NULL;
 362         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
 363 }
 364
 365 static inline void free_mm_slot(struct mm_slot *mm_slot)
 366 {
 367         kmem_cache_free(mm_slot_cache, mm_slot);
 368 }
 369
 370 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
 371 {
 372         struct mm_slot *mm_slot;
 373
 374         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
 375                 if (mm == mm_slot->mm)
 376                         return mm_slot;
 377
 378         return NULL;
 379 }
 380
 381 static void insert_to_mm_slots_hash(struct mm_struct *mm,
 382                                     struct mm_slot *mm_slot)
 383 {
 384         mm_slot->mm = mm;
 385         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
 386 }
 387
 388 static inline int khugepaged_test_exit(struct mm_struct *mm)
 389 {
 390         return atomic_read(&mm->mm_users) == 0;
 391 }
 392
 393 int __khugepaged_enter(struct mm_struct *mm)
 394 {
 395         struct mm_slot *mm_slot;
 396         int wakeup;
 397
 398         mm_slot = alloc_mm_slot();
 399         if (!mm_slot)
 400                 return -ENOMEM;
 401
 402         /* __khugepaged_exit() must not run from under us */
 403         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
 404         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
 405                 free_mm_slot(mm_slot);
 406                 return 0;
 407         }
 408
 409         spin_lock(&khugepaged_mm_lock);
 410         insert_to_mm_slots_hash(mm, mm_slot);
 411         /*
 412          * Insert just behind the scanning cursor, to let the area settle
 413          * down a little.
 414          */
 415         wakeup = list_empty(&khugepaged_scan.mm_head);
 416         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
 417         spin_unlock(&khugepaged_mm_lock);
 418
 419         atomic_inc(&mm->mm_count);
 420         if (wakeup)
 421                 wake_up_interruptible(&khugepaged_wait);
 422
 423         return 0;
 424 }
 425
 426 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
 427                                unsigned long vm_flags)
 428 {
 429         unsigned long hstart, hend;
 430         if (!vma->anon_vma)
 431                 /*
 432                  * Not yet faulted in so we will register later in the
 433                  * page fault if needed.
 434                  */
 435                 return 0;
 436         if (vma->vm_ops || (vm_flags & VM_NO_KHUGEPAGED))
 437                 /* khugepaged not yet working on file or special mappings */
 438                 return 0;
 439         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 440         hend = vma->vm_end & HPAGE_PMD_MASK;
 441         if (hstart < hend)
 442                 return khugepaged_enter(vma, vm_flags);
 443         return 0;
 444 }
 445
 446 void __khugepaged_exit(struct mm_struct *mm)
 447 {
 448         struct mm_slot *mm_slot;
 449         int free = 0;
 450
 451         spin_lock(&khugepaged_mm_lock);
 452         mm_slot = get_mm_slot(mm);
 453         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
 454                 hash_del(&mm_slot->hash);
 455                 list_del(&mm_slot->mm_node);
 456                 free = 1;
 457         }
 458         spin_unlock(&khugepaged_mm_lock);
 459
 460         if (free) {
 461                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
 462                 free_mm_slot(mm_slot);
 463                 mmdrop(mm);
 464         } else if (mm_slot) {
 465                 /*
 466                  * This is required to serialize against
 467                  * khugepaged_test_exit() (which is guaranteed to run
 468                  * under mmap sem read mode). Stop here (after we
 469                  * return all pagetables will be destroyed) until
 470                  * khugepaged has finished working on the pagetables
 471                  * under the mmap_sem.
 472                  */
 473                 down_write(&mm->mmap_sem);
 474                 up_write(&mm->mmap_sem);
 475         }
 476 }
 477
 478 static void release_pte_page(struct page *page)
 479 {
 480         /* 0 stands for page_is_file_cache(page) == false */
 481         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
 482         unlock_page(page);
 483         putback_lru_page(page);
 484 }
 485
 486 static void release_pte_pages(pte_t *pte, pte_t *_pte)
 487 {
 488         while (--_pte >= pte) {
 489                 pte_t pteval = *_pte;
 490                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
 491                         release_pte_page(pte_page(pteval));
 492         }
 493 }
 494
 495 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
 496                                         unsigned long address,
 497                                         pte_t *pte)
 498 {
 499         struct page *page = NULL;
 500         pte_t *_pte;
 501         int none_or_zero = 0, result = 0;
 502         bool referenced = false, writable = false;
 503
 504         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
 505              _pte++, address += PAGE_SIZE) {
 506                 pte_t pteval = *_pte;
 507                 if (pte_none(pteval) || (pte_present(pteval) &&
 508                                 is_zero_pfn(pte_pfn(pteval)))) {
 509                         if (!userfaultfd_armed(vma) &&
 510                             ++none_or_zero <= khugepaged_max_ptes_none) {
 511                                 continue;
 512                         } else {
 513                                 result = SCAN_EXCEED_NONE_PTE;
 514                                 goto out;
 515                         }
 516                 }
 517                 if (!pte_present(pteval)) {
 518                         result = SCAN_PTE_NON_PRESENT;
 519                         goto out;
 520                 }
 521                 page = vm_normal_page(vma, address, pteval);
 522                 if (unlikely(!page)) {
 523                         result = SCAN_PAGE_NULL;
 524                         goto out;
 525                 }
 526
 527                 VM_BUG_ON_PAGE(PageCompound(page), page);
 528                 VM_BUG_ON_PAGE(!PageAnon(page), page);
 529                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
 530
 531                 /*
 532                  * We can do it before isolate_lru_page because the
 533                  * page can't be freed from under us. NOTE: PG_lock
 534                  * is needed to serialize against split_huge_page
 535                  * when invoked from the VM.
 536                  */
 537                 if (!trylock_page(page)) {
 538                         result = SCAN_PAGE_LOCK;
 539                         goto out;
 540                 }
 541
 542                 /*
 543                  * cannot use mapcount: can't collapse if there's a gup pin.
 544                  * The page must only be referenced by the scanned process
 545                  * and page swap cache.
 546                  */
 547                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
 548                         unlock_page(page);
 549                         result = SCAN_PAGE_COUNT;
 550                         goto out;
 551                 }
 552                 if (pte_write(pteval)) {
 553                         writable = true;
 554                 } else {
 555                         if (PageSwapCache(page) &&
 556                             !reuse_swap_page(page, NULL)) {
 557                                 unlock_page(page);
 558                                 result = SCAN_SWAP_CACHE_PAGE;
 559                                 goto out;
 560                         }
 561                         /*
 562                          * Page is not in the swap cache. It can be collapsed
 563                          * into a THP.
 564                          */
 565                 }
 566
 567                 /*
 568                  * Isolate the page to avoid collapsing an hugepage
 569                  * currently in use by the VM.
 570                  */
 571                 if (isolate_lru_page(page)) {
 572                         unlock_page(page);
 573                         result = SCAN_DEL_PAGE_LRU;
 574                         goto out;
 575                 }
 576                 /* 0 stands for page_is_file_cache(page) == false */
 577                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
 578                 VM_BUG_ON_PAGE(!PageLocked(page), page);
 579                 VM_BUG_ON_PAGE(PageLRU(page), page);
 580
 581                 /* If there is no mapped pte young don't collapse the page */
 582                 if (pte_young(pteval) ||
 583                     page_is_young(page) || PageReferenced(page) ||
 584                     mmu_notifier_test_young(vma->vm_mm, address))
 585                         referenced = true;
 586         }
 587         if (likely(writable)) {
 588                 if (likely(referenced)) {
 589                         result = SCAN_SUCCEED;
 590                         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 591                                                             referenced, writable, result);
 592                         return 1;
 593                 }
 594         } else {
 595                 result = SCAN_PAGE_RO;
 596         }
 597
 598 out:
 599         release_pte_pages(pte, _pte);
 600         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 601                                             referenced, writable, result);
 602         return 0;
 603 }
 604
 605 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
 606                                       struct vm_area_struct *vma,
 607                                       unsigned long address,
 608                                       spinlock_t *ptl)
 609 {
 610         pte_t *_pte;
 611         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
 612                 pte_t pteval = *_pte;
 613                 struct page *src_page;
 614
 615                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
 616                         clear_user_highpage(page, address);
 617                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
 618                         if (is_zero_pfn(pte_pfn(pteval))) {
 619                                 /*
 620                                  * ptl mostly unnecessary.
 621                                  */
 622                                 spin_lock(ptl);
 623                                 /*
 624                                  * paravirt calls inside pte_clear here are
 625                                  * superfluous.
 626                                  */
 627                                 pte_clear(vma->vm_mm, address, _pte);
 628                                 spin_unlock(ptl);
 629                         }
 630                 } else {
 631                         src_page = pte_page(pteval);
 632                         copy_user_highpage(page, src_page, address, vma);
 633                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
 634                         release_pte_page(src_page);
 635                         /*
 636                          * ptl mostly unnecessary, but preempt has to
 637                          * be disabled to update the per-cpu stats
 638                          * inside page_remove_rmap().
 639                          */
 640                         spin_lock(ptl);
 641                         /*
 642                          * paravirt calls inside pte_clear here are
 643                          * superfluous.
 644                          */
 645                         pte_clear(vma->vm_mm, address, _pte);
 646                         page_remove_rmap(src_page, false);
 647                         spin_unlock(ptl);
 648                         free_page_and_swap_cache(src_page);
 649                 }
 650
 651                 address += PAGE_SIZE;
 652                 page++;
 653         }
 654 }
 655
 656 static void khugepaged_alloc_sleep(void)
 657 {
 658         DEFINE_WAIT(wait);
 659
 660         add_wait_queue(&khugepaged_wait, &wait);
 661         freezable_schedule_timeout_interruptible(
 662                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
 663         remove_wait_queue(&khugepaged_wait, &wait);
 664 }
 665
 666 static int khugepaged_node_load[MAX_NUMNODES];
 667
 668 static bool khugepaged_scan_abort(int nid)
 669 {
 670         int i;
 671
 672         /*
 673          * If zone_reclaim_mode is disabled, then no extra effort is made to
 674          * allocate memory locally.
 675          */
 676         if (!zone_reclaim_mode)
 677                 return false;
 678
 679         /* If there is a count for this node already, it must be acceptable */
 680         if (khugepaged_node_load[nid])
 681                 return false;
 682
 683         for (i = 0; i < MAX_NUMNODES; i++) {
 684                 if (!khugepaged_node_load[i])
 685                         continue;
 686                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
 687                         return true;
 688         }
 689         return false;
 690 }
 691
 692 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
 693 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
 694 {
 695         return GFP_TRANSHUGE | (khugepaged_defrag() ? __GFP_DIRECT_RECLAIM : 0);
 696 }
 697
 698 #ifdef CONFIG_NUMA
 699 static int khugepaged_find_target_node(void)
 700 {
 701         static int last_khugepaged_target_node = NUMA_NO_NODE;
 702         int nid, target_node = 0, max_value = 0;
 703
 704         /* find first node with max normal pages hit */
 705         for (nid = 0; nid < MAX_NUMNODES; nid++)
 706                 if (khugepaged_node_load[nid] > max_value) {
 707                         max_value = khugepaged_node_load[nid];
 708                         target_node = nid;
 709                 }
 710
 711         /* do some balance if several nodes have the same hit record */
 712         if (target_node <= last_khugepaged_target_node)
 713                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
 714                                 nid++)
 715                         if (max_value == khugepaged_node_load[nid]) {
 716                                 target_node = nid;
 717                                 break;
 718                         }
 719
 720         last_khugepaged_target_node = target_node;
 721         return target_node;
 722 }
 723
 724 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 725 {
 726         if (IS_ERR(*hpage)) {
 727                 if (!*wait)
 728                         return false;
 729
 730                 *wait = false;
 731                 *hpage = NULL;
 732                 khugepaged_alloc_sleep();
 733         } else if (*hpage) {
 734                 put_page(*hpage);
 735                 *hpage = NULL;
 736         }
 737
 738         return true;
 739 }
 740
 741 static struct page *
 742 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
 743                        unsigned long address, int node)
 744 {
 745         VM_BUG_ON_PAGE(*hpage, *hpage);
 746
 747         /*
 748          * Before allocating the hugepage, release the mmap_sem read lock.
 749          * The allocation can take potentially a long time if it involves
 750          * sync compaction, and we do not need to hold the mmap_sem during
 751          * that. We will recheck the vma after taking it again in write mode.
 752          */
 753         up_read(&mm->mmap_sem);
 754
 755         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
 756         if (unlikely(!*hpage)) {
 757                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 758                 *hpage = ERR_PTR(-ENOMEM);
 759                 return NULL;
 760         }
 761
 762         prep_transhuge_page(*hpage);
 763         count_vm_event(THP_COLLAPSE_ALLOC);
 764         return *hpage;
 765 }
 766 #else
 767 static int khugepaged_find_target_node(void)
 768 {
 769         return 0;
 770 }
 771
 772 static inline struct page *alloc_khugepaged_hugepage(void)
 773 {
 774         struct page *page;
 775
 776         page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
 777                            HPAGE_PMD_ORDER);
 778         if (page)
 779                 prep_transhuge_page(page);
 780         return page;
 781 }
 782
 783 static struct page *khugepaged_alloc_hugepage(bool *wait)
 784 {
 785         struct page *hpage;
 786
 787         do {
 788                 hpage = alloc_khugepaged_hugepage();
 789                 if (!hpage) {
 790                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 791                         if (!*wait)
 792                                 return NULL;
 793
 794                         *wait = false;
 795                         khugepaged_alloc_sleep();
 796                 } else
 797                         count_vm_event(THP_COLLAPSE_ALLOC);
 798         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
 799
 800         return hpage;
 801 }
 802
 803 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 804 {
 805         if (!*hpage)
 806                 *hpage = khugepaged_alloc_hugepage(wait);
 807
 808         if (unlikely(!*hpage))
 809                 return false;
 810
 811         return true;
 812 }
 813
 814 static struct page *
 815 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
 816                        unsigned long address, int node)
 817 {
 818         up_read(&mm->mmap_sem);
 819         VM_BUG_ON(!*hpage);
 820
 821         return  *hpage;
 822 }
 823 #endif
 824
 825 static bool hugepage_vma_check(struct vm_area_struct *vma)
 826 {
 827         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
 828             (vma->vm_flags & VM_NOHUGEPAGE))
 829                 return false;
 830         if (!vma->anon_vma || vma->vm_ops)
 831                 return false;
 832         if (is_vma_temporary_stack(vma))
 833                 return false;
 834         return !(vma->vm_flags & VM_NO_KHUGEPAGED);
 835 }
 836
 837 /*
 838  * If mmap_sem temporarily dropped, revalidate vma
 839  * before taking mmap_sem.
 840  * Return 0 if succeeds, otherwise return none-zero
 841  * value (scan code).
 842  */
 843
 844 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address)
 845 {
 846         struct vm_area_struct *vma;
 847         unsigned long hstart, hend;
 848
 849         if (unlikely(khugepaged_test_exit(mm)))
 850                 return SCAN_ANY_PROCESS;
 851
 852         vma = find_vma(mm, address);
 853         if (!vma)
 854                 return SCAN_VMA_NULL;
 855
 856         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 857         hend = vma->vm_end & HPAGE_PMD_MASK;
 858         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
 859                 return SCAN_ADDRESS_RANGE;
 860         if (!hugepage_vma_check(vma))
 861                 return SCAN_VMA_CHECK;
 862         return 0;
 863 }
 864
 865 /*
 866  * Bring missing pages in from swap, to complete THP collapse.
 867  * Only done if khugepaged_scan_pmd believes it is worthwhile.
 868  *
 869  * Called and returns without pte mapped or spinlocks held,
 870  * but with mmap_sem held to protect against vma changes.
 871  */
 872
 873 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
 874                                         struct vm_area_struct *vma,
 875                                         unsigned long address, pmd_t *pmd)
 876 {
 877         pte_t pteval;
 878         int swapped_in = 0, ret = 0;
 879         struct fault_env fe = {
 880                 .vma = vma,
 881                 .address = address,
 882                 .flags = FAULT_FLAG_ALLOW_RETRY,
 883                 .pmd = pmd,
 884         };
 885
 886         fe.pte = pte_offset_map(pmd, address);
 887         for (; fe.address < address + HPAGE_PMD_NR*PAGE_SIZE;
 888                         fe.pte++, fe.address += PAGE_SIZE) {
 889                 pteval = *fe.pte;
 890                 if (!is_swap_pte(pteval))
 891                         continue;
 892                 swapped_in++;
 893                 ret = do_swap_page(&fe, pteval);
 894                 /* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */
 895                 if (ret & VM_FAULT_RETRY) {
 896                         down_read(&mm->mmap_sem);
 897                         /* vma is no longer available, don't continue to swapin */
 898                         if (hugepage_vma_revalidate(mm, address))
 899                                 return false;
 900                         /* check if the pmd is still valid */
 901                         if (mm_find_pmd(mm, address) != pmd)
 902                                 return false;
 903                 }
 904                 if (ret & VM_FAULT_ERROR) {
 905                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, 0);
 906                         return false;
 907                 }
 908                 /* pte is unmapped now, we need to map it */
 909                 fe.pte = pte_offset_map(pmd, fe.address);
 910         }
 911         fe.pte--;
 912         pte_unmap(fe.pte);
 913         trace_mm_collapse_huge_page_swapin(mm, swapped_in, 1);
 914         return true;
 915 }
 916
 917 static void collapse_huge_page(struct mm_struct *mm,
 918                                    unsigned long address,
 919                                    struct page **hpage,
 920                                    struct vm_area_struct *vma,
 921                                    int node)
 922 {
 923         pmd_t *pmd, _pmd;
 924         pte_t *pte;
 925         pgtable_t pgtable;
 926         struct page *new_page;
 927         spinlock_t *pmd_ptl, *pte_ptl;
 928         int isolated = 0, result = 0;
 929         struct mem_cgroup *memcg;
 930         unsigned long mmun_start;       /* For mmu_notifiers */
 931         unsigned long mmun_end;         /* For mmu_notifiers */
 932         gfp_t gfp;
 933
 934         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 935
 936         /* Only allocate from the target node */
 937         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_OTHER_NODE | __GFP_THISNODE;
 938
 939         /* release the mmap_sem read lock. */
 940         new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node);
 941         if (!new_page) {
 942                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
 943                 goto out_nolock;
 944         }
 945
 946         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
 947                 result = SCAN_CGROUP_CHARGE_FAIL;
 948                 goto out_nolock;
 949         }
 950
 951         down_read(&mm->mmap_sem);
 952         result = hugepage_vma_revalidate(mm, address);
 953         if (result) {
 954                 mem_cgroup_cancel_charge(new_page, memcg, true);
 955                 up_read(&mm->mmap_sem);
 956                 goto out_nolock;
 957         }
 958
 959         pmd = mm_find_pmd(mm, address);
 960         if (!pmd) {
 961                 result = SCAN_PMD_NULL;
 962                 mem_cgroup_cancel_charge(new_page, memcg, true);
 963                 up_read(&mm->mmap_sem);
 964                 goto out_nolock;
 965         }
 966
 967         /*
 968          * __collapse_huge_page_swapin always returns with mmap_sem locked.
 969          * If it fails, release mmap_sem and jump directly out.
 970          * Continuing to collapse causes inconsistency.
 971          */
 972         if (!__collapse_huge_page_swapin(mm, vma, address, pmd)) {
 973                 mem_cgroup_cancel_charge(new_page, memcg, true);
 974                 up_read(&mm->mmap_sem);
 975                 goto out_nolock;
 976         }
 977
 978         up_read(&mm->mmap_sem);
 979         /*
 980          * Prevent all access to pagetables with the exception of
 981          * gup_fast later handled by the ptep_clear_flush and the VM
 982          * handled by the anon_vma lock + PG_lock.
 983          */
 984         down_write(&mm->mmap_sem);
 985         result = hugepage_vma_revalidate(mm, address);
 986         if (result)
 987                 goto out;
 988         /* check if the pmd is still valid */
 989         if (mm_find_pmd(mm, address) != pmd)
 990                 goto out;
 991
 992         anon_vma_lock_write(vma->anon_vma);
 993
 994         pte = pte_offset_map(pmd, address);
 995         pte_ptl = pte_lockptr(mm, pmd);
 996
 997         mmun_start = address;
 998         mmun_end   = address + HPAGE_PMD_SIZE;
 999         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1000         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1001         /*
1002          * After this gup_fast can't run anymore. This also removes
1003          * any huge TLB entry from the CPU so we won't allow
1004          * huge and small TLB entries for the same virtual address
1005          * to avoid the risk of CPU bugs in that area.
1006          */
1007         _pmd = pmdp_collapse_flush(vma, address, pmd);
1008         spin_unlock(pmd_ptl);
1009         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1010
1011         spin_lock(pte_ptl);
1012         isolated = __collapse_huge_page_isolate(vma, address, pte);
1013         spin_unlock(pte_ptl);
1014
1015         if (unlikely(!isolated)) {
1016                 pte_unmap(pte);
1017                 spin_lock(pmd_ptl);
1018                 BUG_ON(!pmd_none(*pmd));
1019                 /*
1020                  * We can only use set_pmd_at when establishing
1021                  * hugepmds and never for establishing regular pmds that
1022                  * points to regular pagetables. Use pmd_populate for that
1023                  */
1024                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1025                 spin_unlock(pmd_ptl);
1026                 anon_vma_unlock_write(vma->anon_vma);
1027                 result = SCAN_FAIL;
1028                 goto out;
1029         }
1030
1031         /*
1032          * All pages are isolated and locked so anon_vma rmap
1033          * can't run anymore.
1034          */
1035         anon_vma_unlock_write(vma->anon_vma);
1036
1037         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
1038         pte_unmap(pte);
1039         __SetPageUptodate(new_page);
1040         pgtable = pmd_pgtable(_pmd);
1041
1042         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1043         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1044
1045         /*
1046          * spin_lock() below is not the equivalent of smp_wmb(), so
1047          * this is needed to avoid the copy_huge_page writes to become
1048          * visible after the set_pmd_at() write.
1049          */
1050         smp_wmb();
1051
1052         spin_lock(pmd_ptl);
1053         BUG_ON(!pmd_none(*pmd));
1054         page_add_new_anon_rmap(new_page, vma, address, true);
1055         mem_cgroup_commit_charge(new_page, memcg, false, true);
1056         lru_cache_add_active_or_unevictable(new_page, vma);
1057         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1058         set_pmd_at(mm, address, pmd, _pmd);
1059         update_mmu_cache_pmd(vma, address, pmd);
1060         spin_unlock(pmd_ptl);
1061
1062         *hpage = NULL;
1063
1064         khugepaged_pages_collapsed++;
1065         result = SCAN_SUCCEED;
1066 out_up_write:
1067         up_write(&mm->mmap_sem);
1068 out_nolock:
1069         trace_mm_collapse_huge_page(mm, isolated, result);
1070         return;
1071 out:
1072         mem_cgroup_cancel_charge(new_page, memcg, true);
1073         goto out_up_write;
1074 }
1075
1076 static int khugepaged_scan_pmd(struct mm_struct *mm,
1077                                struct vm_area_struct *vma,
1078                                unsigned long address,
1079                                struct page **hpage)
1080 {
1081         pmd_t *pmd;
1082         pte_t *pte, *_pte;
1083         int ret = 0, none_or_zero = 0, result = 0;
1084         struct page *page = NULL;
1085         unsigned long _address;
1086         spinlock_t *ptl;
1087         int node = NUMA_NO_NODE, unmapped = 0;
1088         bool writable = false, referenced = false;
1089
1090         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1091
1092         pmd = mm_find_pmd(mm, address);
1093         if (!pmd) {
1094                 result = SCAN_PMD_NULL;
1095                 goto out;
1096         }
1097
1098         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1099         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1100         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1101              _pte++, _address += PAGE_SIZE) {
1102                 pte_t pteval = *_pte;
1103                 if (is_swap_pte(pteval)) {
1104                         if (++unmapped <= khugepaged_max_ptes_swap) {
1105                                 continue;
1106                         } else {
1107                                 result = SCAN_EXCEED_SWAP_PTE;
1108                                 goto out_unmap;
1109                         }
1110                 }
1111                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1112                         if (!userfaultfd_armed(vma) &&
1113                             ++none_or_zero <= khugepaged_max_ptes_none) {
1114                                 continue;
1115                         } else {
1116                                 result = SCAN_EXCEED_NONE_PTE;
1117                                 goto out_unmap;
1118                         }
1119                 }
1120                 if (!pte_present(pteval)) {
1121                         result = SCAN_PTE_NON_PRESENT;
1122                         goto out_unmap;
1123                 }
1124                 if (pte_write(pteval))
1125                         writable = true;
1126
1127                 page = vm_normal_page(vma, _address, pteval);
1128                 if (unlikely(!page)) {
1129                         result = SCAN_PAGE_NULL;
1130                         goto out_unmap;
1131                 }
1132
1133                 /* TODO: teach khugepaged to collapse THP mapped with pte */
1134                 if (PageCompound(page)) {
1135                         result = SCAN_PAGE_COMPOUND;
1136                         goto out_unmap;
1137                 }
1138
1139                 /*
1140                  * Record which node the original page is from and save this
1141                  * information to khugepaged_node_load[].
1142                  * Khupaged will allocate hugepage from the node has the max
1143                  * hit record.
1144                  */
1145                 node = page_to_nid(page);
1146                 if (khugepaged_scan_abort(node)) {
1147                         result = SCAN_SCAN_ABORT;
1148                         goto out_unmap;
1149                 }
1150                 khugepaged_node_load[node]++;
1151                 if (!PageLRU(page)) {
1152                         result = SCAN_PAGE_LRU;
1153                         goto out_unmap;
1154                 }
1155                 if (PageLocked(page)) {
1156                         result = SCAN_PAGE_LOCK;
1157                         goto out_unmap;
1158                 }
1159                 if (!PageAnon(page)) {
1160                         result = SCAN_PAGE_ANON;
1161                         goto out_unmap;
1162                 }
1163
1164                 /*
1165                  * cannot use mapcount: can't collapse if there's a gup pin.
1166                  * The page must only be referenced by the scanned process
1167                  * and page swap cache.
1168                  */
1169                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
1170                         result = SCAN_PAGE_COUNT;
1171                         goto out_unmap;
1172                 }
1173                 if (pte_young(pteval) ||
1174                     page_is_young(page) || PageReferenced(page) ||
1175                     mmu_notifier_test_young(vma->vm_mm, address))
1176                         referenced = true;
1177         }
1178         if (writable) {
1179                 if (referenced) {
1180                         result = SCAN_SUCCEED;
1181                         ret = 1;
1182                 } else {
1183                         result = SCAN_NO_REFERENCED_PAGE;
1184                 }
1185         } else {
1186                 result = SCAN_PAGE_RO;
1187         }
1188 out_unmap:
1189         pte_unmap_unlock(pte, ptl);
1190         if (ret) {
1191                 node = khugepaged_find_target_node();
1192                 /* collapse_huge_page will return with the mmap_sem released */
1193                 collapse_huge_page(mm, address, hpage, vma, node);
1194         }
1195 out:
1196         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1197                                      none_or_zero, result, unmapped);
1198         return ret;
1199 }
1200
1201 static void collect_mm_slot(struct mm_slot *mm_slot)
1202 {
1203         struct mm_struct *mm = mm_slot->mm;
1204
1205         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
1206
1207         if (khugepaged_test_exit(mm)) {
1208                 /* free mm_slot */
1209                 hash_del(&mm_slot->hash);
1210                 list_del(&mm_slot->mm_node);
1211
1212                 /*
1213                  * Not strictly needed because the mm exited already.
1214                  *
1215                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1216                  */
1217
1218                 /* khugepaged_mm_lock actually not necessary for the below */
1219                 free_mm_slot(mm_slot);
1220                 mmdrop(mm);
1221         }
1222 }
1223
1224 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
1225                                             struct page **hpage)
1226         __releases(&khugepaged_mm_lock)
1227         __acquires(&khugepaged_mm_lock)
1228 {
1229         struct mm_slot *mm_slot;
1230         struct mm_struct *mm;
1231         struct vm_area_struct *vma;
1232         int progress = 0;
1233
1234         VM_BUG_ON(!pages);
1235         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
1236
1237         if (khugepaged_scan.mm_slot)
1238                 mm_slot = khugepaged_scan.mm_slot;
1239         else {
1240                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
1241                                      struct mm_slot, mm_node);
1242                 khugepaged_scan.address = 0;
1243                 khugepaged_scan.mm_slot = mm_slot;
1244         }
1245         spin_unlock(&khugepaged_mm_lock);
1246
1247         mm = mm_slot->mm;
1248         down_read(&mm->mmap_sem);
1249         if (unlikely(khugepaged_test_exit(mm)))
1250                 vma = NULL;
1251         else
1252                 vma = find_vma(mm, khugepaged_scan.address);
1253
1254         progress++;
1255         for (; vma; vma = vma->vm_next) {
1256                 unsigned long hstart, hend;
1257
1258                 cond_resched();
1259                 if (unlikely(khugepaged_test_exit(mm))) {
1260                         progress++;
1261                         break;
1262                 }
1263                 if (!hugepage_vma_check(vma)) {
1264 skip:
1265                         progress++;
1266                         continue;
1267                 }
1268                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1269                 hend = vma->vm_end & HPAGE_PMD_MASK;
1270                 if (hstart >= hend)
1271                         goto skip;
1272                 if (khugepaged_scan.address > hend)
1273                         goto skip;
1274                 if (khugepaged_scan.address < hstart)
1275                         khugepaged_scan.address = hstart;
1276                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
1277
1278                 while (khugepaged_scan.address < hend) {
1279                         int ret;
1280                         cond_resched();
1281                         if (unlikely(khugepaged_test_exit(mm)))
1282                                 goto breakouterloop;
1283
1284                         VM_BUG_ON(khugepaged_scan.address < hstart ||
1285                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
1286                                   hend);
1287                         ret = khugepaged_scan_pmd(mm, vma,
1288                                                   khugepaged_scan.address,
1289                                                   hpage);
1290                         /* move to next address */
1291                         khugepaged_scan.address += HPAGE_PMD_SIZE;
1292                         progress += HPAGE_PMD_NR;
1293                         if (ret)
1294                                 /* we released mmap_sem so break loop */
1295                                 goto breakouterloop_mmap_sem;
1296                         if (progress >= pages)
1297                                 goto breakouterloop;
1298                 }
1299         }
1300 breakouterloop:
1301         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
1302 breakouterloop_mmap_sem:
1303
1304         spin_lock(&khugepaged_mm_lock);
1305         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
1306         /*
1307          * Release the current mm_slot if this mm is about to die, or
1308          * if we scanned all vmas of this mm.
1309          */
1310         if (khugepaged_test_exit(mm) || !vma) {
1311                 /*
1312                  * Make sure that if mm_users is reaching zero while
1313                  * khugepaged runs here, khugepaged_exit will find
1314                  * mm_slot not pointing to the exiting mm.
1315                  */
1316                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
1317                         khugepaged_scan.mm_slot = list_entry(
1318                                 mm_slot->mm_node.next,
1319                                 struct mm_slot, mm_node);
1320                         khugepaged_scan.address = 0;
1321                 } else {
1322                         khugepaged_scan.mm_slot = NULL;
1323                         khugepaged_full_scans++;
1324                 }
1325
1326                 collect_mm_slot(mm_slot);
1327         }
1328
1329         return progress;
1330 }
1331
1332 static int khugepaged_has_work(void)
1333 {
1334         return !list_empty(&khugepaged_scan.mm_head) &&
1335                 khugepaged_enabled();
1336 }
1337
1338 static int khugepaged_wait_event(void)
1339 {
1340         return !list_empty(&khugepaged_scan.mm_head) ||
1341                 kthread_should_stop();
1342 }
1343
1344 static void khugepaged_do_scan(void)
1345 {
1346         struct page *hpage = NULL;
1347         unsigned int progress = 0, pass_through_head = 0;
1348         unsigned int pages = khugepaged_pages_to_scan;
1349         bool wait = true;
1350
1351         barrier(); /* write khugepaged_pages_to_scan to local stack */
1352
1353         while (progress < pages) {
1354                 if (!khugepaged_prealloc_page(&hpage, &wait))
1355                         break;
1356
1357                 cond_resched();
1358
1359                 if (unlikely(kthread_should_stop() || try_to_freeze()))
1360                         break;
1361
1362                 spin_lock(&khugepaged_mm_lock);
1363                 if (!khugepaged_scan.mm_slot)
1364                         pass_through_head++;
1365                 if (khugepaged_has_work() &&
1366                     pass_through_head < 2)
1367                         progress += khugepaged_scan_mm_slot(pages - progress,
1368                                                             &hpage);
1369                 else
1370                         progress = pages;
1371                 spin_unlock(&khugepaged_mm_lock);
1372         }
1373
1374         if (!IS_ERR_OR_NULL(hpage))
1375                 put_page(hpage);
1376 }
1377
1378 static bool khugepaged_should_wakeup(void)
1379 {
1380         return kthread_should_stop() ||
1381                time_after_eq(jiffies, khugepaged_sleep_expire);
1382 }
1383
1384 static void khugepaged_wait_work(void)
1385 {
1386         if (khugepaged_has_work()) {
1387                 const unsigned long scan_sleep_jiffies =
1388                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
1389
1390                 if (!scan_sleep_jiffies)
1391                         return;
1392
1393                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
1394                 wait_event_freezable_timeout(khugepaged_wait,
1395                                              khugepaged_should_wakeup(),
1396                                              scan_sleep_jiffies);
1397                 return;
1398         }
1399
1400         if (khugepaged_enabled())
1401                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
1402 }
1403
1404 static int khugepaged(void *none)
1405 {
1406         struct mm_slot *mm_slot;
1407
1408         set_freezable();
1409         set_user_nice(current, MAX_NICE);
1410
1411         while (!kthread_should_stop()) {
1412                 khugepaged_do_scan();
1413                 khugepaged_wait_work();
1414         }
1415
1416         spin_lock(&khugepaged_mm_lock);
1417         mm_slot = khugepaged_scan.mm_slot;
1418         khugepaged_scan.mm_slot = NULL;
1419         if (mm_slot)
1420                 collect_mm_slot(mm_slot);
1421         spin_unlock(&khugepaged_mm_lock);
1422         return 0;
1423 }
1424
1425 static void set_recommended_min_free_kbytes(void)
1426 {
1427         struct zone *zone;
1428         int nr_zones = 0;
1429         unsigned long recommended_min;
1430
1431         for_each_populated_zone(zone)
1432                 nr_zones++;
1433
1434         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
1435         recommended_min = pageblock_nr_pages * nr_zones * 2;
1436
1437         /*
1438          * Make sure that on average at least two pageblocks are almost free
1439          * of another type, one for a migratetype to fall back to and a
1440          * second to avoid subsequent fallbacks of other types There are 3
1441          * MIGRATE_TYPES we care about.
1442          */
1443         recommended_min += pageblock_nr_pages * nr_zones *
1444                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
1445
1446         /* don't ever allow to reserve more than 5% of the lowmem */
1447         recommended_min = min(recommended_min,
1448                               (unsigned long) nr_free_buffer_pages() / 20);
1449         recommended_min <<= (PAGE_SHIFT-10);
1450
1451         if (recommended_min > min_free_kbytes) {
1452                 if (user_min_free_kbytes >= 0)
1453                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
1454                                 min_free_kbytes, recommended_min);
1455
1456                 min_free_kbytes = recommended_min;
1457         }
1458         setup_per_zone_wmarks();
1459 }
1460
1461 int start_stop_khugepaged(void)
1462 {
1463         static struct task_struct *khugepaged_thread __read_mostly;
1464         static DEFINE_MUTEX(khugepaged_mutex);
1465         int err = 0;
1466
1467         mutex_lock(&khugepaged_mutex);
1468         if (khugepaged_enabled()) {
1469                 if (!khugepaged_thread)
1470                         khugepaged_thread = kthread_run(khugepaged, NULL,
1471                                                         "khugepaged");
1472                 if (IS_ERR(khugepaged_thread)) {
1473                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
1474                         err = PTR_ERR(khugepaged_thread);
1475                         khugepaged_thread = NULL;
1476                         goto fail;
1477                 }
1478
1479                 if (!list_empty(&khugepaged_scan.mm_head))
1480                         wake_up_interruptible(&khugepaged_wait);
1481
1482                 set_recommended_min_free_kbytes();
1483         } else if (khugepaged_thread) {
1484                 kthread_stop(khugepaged_thread);
1485                 khugepaged_thread = NULL;
1486         }
1487 fail:
1488         mutex_unlock(&khugepaged_mutex);
1489         return err;
1490 }