mm/khugepaged.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   3
   4 #include <linux/mm.h>
   5 #include <linux/sched.h>
   6 #include <linux/sched/mm.h>
   7 #include <linux/sched/coredump.h>
   8 #include <linux/mmu_notifier.h>
   9 #include <linux/rmap.h>
  10 #include <linux/swap.h>
  11 #include <linux/mm_inline.h>
  12 #include <linux/kthread.h>
  13 #include <linux/khugepaged.h>
  14 #include <linux/freezer.h>
  15 #include <linux/mman.h>
  16 #include <linux/hashtable.h>
  17 #include <linux/userfaultfd_k.h>
  18 #include <linux/page_idle.h>
  19 #include <linux/swapops.h>
  20 #include <linux/shmem_fs.h>
  21
  22 #include <asm/tlb.h>
  23 #include <asm/pgalloc.h>
  24 #include "internal.h"
  25
  26 enum scan_result {
  27         SCAN_FAIL,
  28         SCAN_SUCCEED,
  29         SCAN_PMD_NULL,
  30         SCAN_EXCEED_NONE_PTE,
  31         SCAN_PTE_NON_PRESENT,
  32         SCAN_PAGE_RO,
  33         SCAN_LACK_REFERENCED_PAGE,
  34         SCAN_PAGE_NULL,
  35         SCAN_SCAN_ABORT,
  36         SCAN_PAGE_COUNT,
  37         SCAN_PAGE_LRU,
  38         SCAN_PAGE_LOCK,
  39         SCAN_PAGE_ANON,
  40         SCAN_PAGE_COMPOUND,
  41         SCAN_ANY_PROCESS,
  42         SCAN_VMA_NULL,
  43         SCAN_VMA_CHECK,
  44         SCAN_ADDRESS_RANGE,
  45         SCAN_SWAP_CACHE_PAGE,
  46         SCAN_DEL_PAGE_LRU,
  47         SCAN_ALLOC_HUGE_PAGE_FAIL,
  48         SCAN_CGROUP_CHARGE_FAIL,
  49         SCAN_EXCEED_SWAP_PTE,
  50         SCAN_TRUNCATED,
  51 };
  52
  53 #define CREATE_TRACE_POINTS
  54 #include <trace/events/huge_memory.h>
  55
  56 static struct task_struct *khugepaged_thread __read_mostly;
  57 static DEFINE_MUTEX(khugepaged_mutex);
  58
  59 /* default scan 8*512 pte (or vmas) every 30 second */
  60 static unsigned int khugepaged_pages_to_scan __read_mostly;
  61 static unsigned int khugepaged_pages_collapsed;
  62 static unsigned int khugepaged_full_scans;
  63 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  64 /* during fragmentation poll the hugepage allocator once every minute */
  65 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  66 static unsigned long khugepaged_sleep_expire;
  67 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  68 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  69 /*
  70  * default collapse hugepages if there is at least one pte mapped like
  71  * it would have happened if the vma was large enough during page
  72  * fault.
  73  */
  74 static unsigned int khugepaged_max_ptes_none __read_mostly;
  75 static unsigned int khugepaged_max_ptes_swap __read_mostly;
  76
  77 #define MM_SLOTS_HASH_BITS 10
  78 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  79
  80 static struct kmem_cache *mm_slot_cache __read_mostly;
  81
  82 /**
  83  * struct mm_slot - hash lookup from mm to mm_slot
  84  * @hash: hash collision list
  85  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  86  * @mm: the mm that this information is valid for
  87  */
  88 struct mm_slot {
  89         struct hlist_node hash;
  90         struct list_head mm_node;
  91         struct mm_struct *mm;
  92 };
  93
  94 /**
  95  * struct khugepaged_scan - cursor for scanning
  96  * @mm_head: the head of the mm list to scan
  97  * @mm_slot: the current mm_slot we are scanning
  98  * @address: the next address inside that to be scanned
  99  *
 100  * There is only the one khugepaged_scan instance of this cursor structure.
 101  */
 102 struct khugepaged_scan {
 103         struct list_head mm_head;
 104         struct mm_slot *mm_slot;
 105         unsigned long address;
 106 };
 107
 108 static struct khugepaged_scan khugepaged_scan = {
 109         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 110 };
 111
 112 #ifdef CONFIG_SYSFS
 113 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 114                                          struct kobj_attribute *attr,
 115                                          char *buf)
 116 {
 117         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 118 }
 119
 120 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 121                                           struct kobj_attribute *attr,
 122                                           const char *buf, size_t count)
 123 {
 124         unsigned long msecs;
 125         int err;
 126
 127         err = kstrtoul(buf, 10, &msecs);
 128         if (err || msecs > UINT_MAX)
 129                 return -EINVAL;
 130
 131         khugepaged_scan_sleep_millisecs = msecs;
 132         khugepaged_sleep_expire = 0;
 133         wake_up_interruptible(&khugepaged_wait);
 134
 135         return count;
 136 }
 137 static struct kobj_attribute scan_sleep_millisecs_attr =
 138         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 139                scan_sleep_millisecs_store);
 140
 141 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 142                                           struct kobj_attribute *attr,
 143                                           char *buf)
 144 {
 145         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 146 }
 147
 148 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 149                                            struct kobj_attribute *attr,
 150                                            const char *buf, size_t count)
 151 {
 152         unsigned long msecs;
 153         int err;
 154
 155         err = kstrtoul(buf, 10, &msecs);
 156         if (err || msecs > UINT_MAX)
 157                 return -EINVAL;
 158
 159         khugepaged_alloc_sleep_millisecs = msecs;
 160         khugepaged_sleep_expire = 0;
 161         wake_up_interruptible(&khugepaged_wait);
 162
 163         return count;
 164 }
 165 static struct kobj_attribute alloc_sleep_millisecs_attr =
 166         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 167                alloc_sleep_millisecs_store);
 168
 169 static ssize_t pages_to_scan_show(struct kobject *kobj,
 170                                   struct kobj_attribute *attr,
 171                                   char *buf)
 172 {
 173         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 174 }
 175 static ssize_t pages_to_scan_store(struct kobject *kobj,
 176                                    struct kobj_attribute *attr,
 177                                    const char *buf, size_t count)
 178 {
 179         int err;
 180         unsigned long pages;
 181
 182         err = kstrtoul(buf, 10, &pages);
 183         if (err || !pages || pages > UINT_MAX)
 184                 return -EINVAL;
 185
 186         khugepaged_pages_to_scan = pages;
 187
 188         return count;
 189 }
 190 static struct kobj_attribute pages_to_scan_attr =
 191         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 192                pages_to_scan_store);
 193
 194 static ssize_t pages_collapsed_show(struct kobject *kobj,
 195                                     struct kobj_attribute *attr,
 196                                     char *buf)
 197 {
 198         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 199 }
 200 static struct kobj_attribute pages_collapsed_attr =
 201         __ATTR_RO(pages_collapsed);
 202
 203 static ssize_t full_scans_show(struct kobject *kobj,
 204                                struct kobj_attribute *attr,
 205                                char *buf)
 206 {
 207         return sprintf(buf, "%u\n", khugepaged_full_scans);
 208 }
 209 static struct kobj_attribute full_scans_attr =
 210         __ATTR_RO(full_scans);
 211
 212 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 213                                       struct kobj_attribute *attr, char *buf)
 214 {
 215         return single_hugepage_flag_show(kobj, attr, buf,
 216                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 217 }
 218 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 219                                        struct kobj_attribute *attr,
 220                                        const char *buf, size_t count)
 221 {
 222         return single_hugepage_flag_store(kobj, attr, buf, count,
 223                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 224 }
 225 static struct kobj_attribute khugepaged_defrag_attr =
 226         __ATTR(defrag, 0644, khugepaged_defrag_show,
 227                khugepaged_defrag_store);
 228
 229 /*
 230  * max_ptes_none controls if khugepaged should collapse hugepages over
 231  * any unmapped ptes in turn potentially increasing the memory
 232  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 233  * reduce the available free memory in the system as it
 234  * runs. Increasing max_ptes_none will instead potentially reduce the
 235  * free memory in the system during the khugepaged scan.
 236  */
 237 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 238                                              struct kobj_attribute *attr,
 239                                              char *buf)
 240 {
 241         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 242 }
 243 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 244                                               struct kobj_attribute *attr,
 245                                               const char *buf, size_t count)
 246 {
 247         int err;
 248         unsigned long max_ptes_none;
 249
 250         err = kstrtoul(buf, 10, &max_ptes_none);
 251         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 252                 return -EINVAL;
 253
 254         khugepaged_max_ptes_none = max_ptes_none;
 255
 256         return count;
 257 }
 258 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 259         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 260                khugepaged_max_ptes_none_store);
 261
 262 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
 263                                              struct kobj_attribute *attr,
 264                                              char *buf)
 265 {
 266         return sprintf(buf, "%u\n", khugepaged_max_ptes_swap);
 267 }
 268
 269 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
 270                                               struct kobj_attribute *attr,
 271                                               const char *buf, size_t count)
 272 {
 273         int err;
 274         unsigned long max_ptes_swap;
 275
 276         err  = kstrtoul(buf, 10, &max_ptes_swap);
 277         if (err || max_ptes_swap > HPAGE_PMD_NR-1)
 278                 return -EINVAL;
 279
 280         khugepaged_max_ptes_swap = max_ptes_swap;
 281
 282         return count;
 283 }
 284
 285 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
 286         __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
 287                khugepaged_max_ptes_swap_store);
 288
 289 static struct attribute *khugepaged_attr[] = {
 290         &khugepaged_defrag_attr.attr,
 291         &khugepaged_max_ptes_none_attr.attr,
 292         &pages_to_scan_attr.attr,
 293         &pages_collapsed_attr.attr,
 294         &full_scans_attr.attr,
 295         &scan_sleep_millisecs_attr.attr,
 296         &alloc_sleep_millisecs_attr.attr,
 297         &khugepaged_max_ptes_swap_attr.attr,
 298         NULL,
 299 };
 300
 301 struct attribute_group khugepaged_attr_group = {
 302         .attrs = khugepaged_attr,
 303         .name = "khugepaged",
 304 };
 305 #endif /* CONFIG_SYSFS */
 306
 307 #define VM_NO_KHUGEPAGED (VM_SPECIAL | VM_HUGETLB)
 308
 309 int hugepage_madvise(struct vm_area_struct *vma,
 310                      unsigned long *vm_flags, int advice)
 311 {
 312         switch (advice) {
 313         case MADV_HUGEPAGE:
 314 #ifdef CONFIG_S390
 315                 /*
 316                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
 317                  * can't handle this properly after s390_enable_sie, so we simply
 318                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
 319                  */
 320                 if (mm_has_pgste(vma->vm_mm))
 321                         return 0;
 322 #endif
 323                 *vm_flags &= ~VM_NOHUGEPAGE;
 324                 *vm_flags |= VM_HUGEPAGE;
 325                 /*
 326                  * If the vma become good for khugepaged to scan,
 327                  * register it here without waiting a page fault that
 328                  * may not happen any time soon.
 329                  */
 330                 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
 331                                 khugepaged_enter_vma_merge(vma, *vm_flags))
 332                         return -ENOMEM;
 333                 break;
 334         case MADV_NOHUGEPAGE:
 335                 *vm_flags &= ~VM_HUGEPAGE;
 336                 *vm_flags |= VM_NOHUGEPAGE;
 337                 /*
 338                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
 339                  * this vma even if we leave the mm registered in khugepaged if
 340                  * it got registered before VM_NOHUGEPAGE was set.
 341                  */
 342                 break;
 343         }
 344
 345         return 0;
 346 }
 347
 348 int __init khugepaged_init(void)
 349 {
 350         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
 351                                           sizeof(struct mm_slot),
 352                                           __alignof__(struct mm_slot), 0, NULL);
 353         if (!mm_slot_cache)
 354                 return -ENOMEM;
 355
 356         khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
 357         khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
 358         khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
 359
 360         return 0;
 361 }
 362
 363 void __init khugepaged_destroy(void)
 364 {
 365         kmem_cache_destroy(mm_slot_cache);
 366 }
 367
 368 static inline struct mm_slot *alloc_mm_slot(void)
 369 {
 370         if (!mm_slot_cache)     /* initialization failed */
 371                 return NULL;
 372         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
 373 }
 374
 375 static inline void free_mm_slot(struct mm_slot *mm_slot)
 376 {
 377         kmem_cache_free(mm_slot_cache, mm_slot);
 378 }
 379
 380 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
 381 {
 382         struct mm_slot *mm_slot;
 383
 384         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
 385                 if (mm == mm_slot->mm)
 386                         return mm_slot;
 387
 388         return NULL;
 389 }
 390
 391 static void insert_to_mm_slots_hash(struct mm_struct *mm,
 392                                     struct mm_slot *mm_slot)
 393 {
 394         mm_slot->mm = mm;
 395         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
 396 }
 397
 398 static inline int khugepaged_test_exit(struct mm_struct *mm)
 399 {
 400         return atomic_read(&mm->mm_users) == 0 || !mmget_still_valid(mm);
 401 }
 402
 403 static bool hugepage_vma_check(struct vm_area_struct *vma,
 404                                unsigned long vm_flags)
 405 {
 406         if ((!(vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
 407             (vm_flags & VM_NOHUGEPAGE) ||
 408             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
 409                 return false;
 410         if (shmem_file(vma->vm_file)) {
 411                 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
 412                         return false;
 413                 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
 414                                 HPAGE_PMD_NR);
 415         }
 416         if (!vma->anon_vma || vma->vm_ops)
 417                 return false;
 418         if (is_vma_temporary_stack(vma))
 419                 return false;
 420         return !(vm_flags & VM_NO_KHUGEPAGED);
 421 }
 422
 423 int __khugepaged_enter(struct mm_struct *mm)
 424 {
 425         struct mm_slot *mm_slot;
 426         int wakeup;
 427
 428         mm_slot = alloc_mm_slot();
 429         if (!mm_slot)
 430                 return -ENOMEM;
 431
 432         /* __khugepaged_exit() must not run from under us */
 433         VM_BUG_ON_MM(atomic_read(&mm->mm_users) == 0, mm);
 434         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
 435                 free_mm_slot(mm_slot);
 436                 return 0;
 437         }
 438
 439         spin_lock(&khugepaged_mm_lock);
 440         insert_to_mm_slots_hash(mm, mm_slot);
 441         /*
 442          * Insert just behind the scanning cursor, to let the area settle
 443          * down a little.
 444          */
 445         wakeup = list_empty(&khugepaged_scan.mm_head);
 446         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
 447         spin_unlock(&khugepaged_mm_lock);
 448
 449         mmgrab(mm);
 450         if (wakeup)
 451                 wake_up_interruptible(&khugepaged_wait);
 452
 453         return 0;
 454 }
 455
 456 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
 457                                unsigned long vm_flags)
 458 {
 459         unsigned long hstart, hend;
 460
 461         /*
 462          * khugepaged does not yet work on non-shmem files or special
 463          * mappings. And file-private shmem THP is not supported.
 464          */
 465         if (!hugepage_vma_check(vma, vm_flags))
 466                 return 0;
 467
 468         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 469         hend = vma->vm_end & HPAGE_PMD_MASK;
 470         if (hstart < hend)
 471                 return khugepaged_enter(vma, vm_flags);
 472         return 0;
 473 }
 474
 475 void __khugepaged_exit(struct mm_struct *mm)
 476 {
 477         struct mm_slot *mm_slot;
 478         int free = 0;
 479
 480         spin_lock(&khugepaged_mm_lock);
 481         mm_slot = get_mm_slot(mm);
 482         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
 483                 hash_del(&mm_slot->hash);
 484                 list_del(&mm_slot->mm_node);
 485                 free = 1;
 486         }
 487         spin_unlock(&khugepaged_mm_lock);
 488
 489         if (free) {
 490                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
 491                 free_mm_slot(mm_slot);
 492                 mmdrop(mm);
 493         } else if (mm_slot) {
 494                 /*
 495                  * This is required to serialize against
 496                  * khugepaged_test_exit() (which is guaranteed to run
 497                  * under mmap sem read mode). Stop here (after we
 498                  * return all pagetables will be destroyed) until
 499                  * khugepaged has finished working on the pagetables
 500                  * under the mmap_sem.
 501                  */
 502                 down_write(&mm->mmap_sem);
 503                 up_write(&mm->mmap_sem);
 504         }
 505 }
 506
 507 static void release_pte_page(struct page *page)
 508 {
 509         dec_node_page_state(page, NR_ISOLATED_ANON + page_is_file_cache(page));
 510         unlock_page(page);
 511         putback_lru_page(page);
 512 }
 513
 514 static void release_pte_pages(pte_t *pte, pte_t *_pte)
 515 {
 516         while (--_pte >= pte) {
 517                 pte_t pteval = *_pte;
 518                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
 519                         release_pte_page(pte_page(pteval));
 520         }
 521 }
 522
 523 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
 524                                         unsigned long address,
 525                                         pte_t *pte)
 526 {
 527         struct page *page = NULL;
 528         pte_t *_pte;
 529         int none_or_zero = 0, result = 0, referenced = 0;
 530         bool writable = false;
 531
 532         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
 533              _pte++, address += PAGE_SIZE) {
 534                 pte_t pteval = *_pte;
 535                 if (pte_none(pteval) || (pte_present(pteval) &&
 536                                 is_zero_pfn(pte_pfn(pteval)))) {
 537                         if (!userfaultfd_armed(vma) &&
 538                             ++none_or_zero <= khugepaged_max_ptes_none) {
 539                                 continue;
 540                         } else {
 541                                 result = SCAN_EXCEED_NONE_PTE;
 542                                 goto out;
 543                         }
 544                 }
 545                 if (!pte_present(pteval)) {
 546                         result = SCAN_PTE_NON_PRESENT;
 547                         goto out;
 548                 }
 549                 page = vm_normal_page(vma, address, pteval);
 550                 if (unlikely(!page)) {
 551                         result = SCAN_PAGE_NULL;
 552                         goto out;
 553                 }
 554
 555                 /* TODO: teach khugepaged to collapse THP mapped with pte */
 556                 if (PageCompound(page)) {
 557                         result = SCAN_PAGE_COMPOUND;
 558                         goto out;
 559                 }
 560
 561                 VM_BUG_ON_PAGE(!PageAnon(page), page);
 562
 563                 /*
 564                  * We can do it before isolate_lru_page because the
 565                  * page can't be freed from under us. NOTE: PG_lock
 566                  * is needed to serialize against split_huge_page
 567                  * when invoked from the VM.
 568                  */
 569                 if (!trylock_page(page)) {
 570                         result = SCAN_PAGE_LOCK;
 571                         goto out;
 572                 }
 573
 574                 /*
 575                  * cannot use mapcount: can't collapse if there's a gup pin.
 576                  * The page must only be referenced by the scanned process
 577                  * and page swap cache.
 578                  */
 579                 if (page_count(page) != 1 + PageSwapCache(page)) {
 580                         unlock_page(page);
 581                         result = SCAN_PAGE_COUNT;
 582                         goto out;
 583                 }
 584                 if (pte_write(pteval)) {
 585                         writable = true;
 586                 } else {
 587                         if (PageSwapCache(page) &&
 588                             !reuse_swap_page(page, NULL)) {
 589                                 unlock_page(page);
 590                                 result = SCAN_SWAP_CACHE_PAGE;
 591                                 goto out;
 592                         }
 593                         /*
 594                          * Page is not in the swap cache. It can be collapsed
 595                          * into a THP.
 596                          */
 597                 }
 598
 599                 /*
 600                  * Isolate the page to avoid collapsing an hugepage
 601                  * currently in use by the VM.
 602                  */
 603                 if (isolate_lru_page(page)) {
 604                         unlock_page(page);
 605                         result = SCAN_DEL_PAGE_LRU;
 606                         goto out;
 607                 }
 608                 inc_node_page_state(page,
 609                                 NR_ISOLATED_ANON + page_is_file_cache(page));
 610                 VM_BUG_ON_PAGE(!PageLocked(page), page);
 611                 VM_BUG_ON_PAGE(PageLRU(page), page);
 612
 613                 /* There should be enough young pte to collapse the page */
 614                 if (pte_young(pteval) ||
 615                     page_is_young(page) || PageReferenced(page) ||
 616                     mmu_notifier_test_young(vma->vm_mm, address))
 617                         referenced++;
 618         }
 619
 620         if (unlikely(!writable)) {
 621                 result = SCAN_PAGE_RO;
 622         } else if (unlikely(!referenced)) {
 623                 result = SCAN_LACK_REFERENCED_PAGE;
 624         } else {
 625                 result = SCAN_SUCCEED;
 626                 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 627                                                     referenced, writable, result);
 628                 return 1;
 629         }
 630 out:
 631         release_pte_pages(pte, _pte);
 632         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 633                                             referenced, writable, result);
 634         return 0;
 635 }
 636
 637 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
 638                                       struct vm_area_struct *vma,
 639                                       unsigned long address,
 640                                       spinlock_t *ptl)
 641 {
 642         pte_t *_pte;
 643         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 644                                 _pte++, page++, address += PAGE_SIZE) {
 645                 pte_t pteval = *_pte;
 646                 struct page *src_page;
 647
 648                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
 649                         clear_user_highpage(page, address);
 650                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
 651                         if (is_zero_pfn(pte_pfn(pteval))) {
 652                                 /*
 653                                  * ptl mostly unnecessary.
 654                                  */
 655                                 spin_lock(ptl);
 656                                 /*
 657                                  * paravirt calls inside pte_clear here are
 658                                  * superfluous.
 659                                  */
 660                                 pte_clear(vma->vm_mm, address, _pte);
 661                                 spin_unlock(ptl);
 662                         }
 663                 } else {
 664                         src_page = pte_page(pteval);
 665                         copy_user_highpage(page, src_page, address, vma);
 666                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
 667                         release_pte_page(src_page);
 668                         /*
 669                          * ptl mostly unnecessary, but preempt has to
 670                          * be disabled to update the per-cpu stats
 671                          * inside page_remove_rmap().
 672                          */
 673                         spin_lock(ptl);
 674                         /*
 675                          * paravirt calls inside pte_clear here are
 676                          * superfluous.
 677                          */
 678                         pte_clear(vma->vm_mm, address, _pte);
 679                         page_remove_rmap(src_page, false);
 680                         spin_unlock(ptl);
 681                         free_page_and_swap_cache(src_page);
 682                 }
 683         }
 684 }
 685
 686 static void khugepaged_alloc_sleep(void)
 687 {
 688         DEFINE_WAIT(wait);
 689
 690         add_wait_queue(&khugepaged_wait, &wait);
 691         freezable_schedule_timeout_interruptible(
 692                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
 693         remove_wait_queue(&khugepaged_wait, &wait);
 694 }
 695
 696 static int khugepaged_node_load[MAX_NUMNODES];
 697
 698 static bool khugepaged_scan_abort(int nid)
 699 {
 700         int i;
 701
 702         /*
 703          * If node_reclaim_mode is disabled, then no extra effort is made to
 704          * allocate memory locally.
 705          */
 706         if (!node_reclaim_mode)
 707                 return false;
 708
 709         /* If there is a count for this node already, it must be acceptable */
 710         if (khugepaged_node_load[nid])
 711                 return false;
 712
 713         for (i = 0; i < MAX_NUMNODES; i++) {
 714                 if (!khugepaged_node_load[i])
 715                         continue;
 716                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
 717                         return true;
 718         }
 719         return false;
 720 }
 721
 722 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
 723 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
 724 {
 725         return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
 726 }
 727
 728 #ifdef CONFIG_NUMA
 729 static int khugepaged_find_target_node(void)
 730 {
 731         static int last_khugepaged_target_node = NUMA_NO_NODE;
 732         int nid, target_node = 0, max_value = 0;
 733
 734         /* find first node with max normal pages hit */
 735         for (nid = 0; nid < MAX_NUMNODES; nid++)
 736                 if (khugepaged_node_load[nid] > max_value) {
 737                         max_value = khugepaged_node_load[nid];
 738                         target_node = nid;
 739                 }
 740
 741         /* do some balance if several nodes have the same hit record */
 742         if (target_node <= last_khugepaged_target_node)
 743                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
 744                                 nid++)
 745                         if (max_value == khugepaged_node_load[nid]) {
 746                                 target_node = nid;
 747                                 break;
 748                         }
 749
 750         last_khugepaged_target_node = target_node;
 751         return target_node;
 752 }
 753
 754 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 755 {
 756         if (IS_ERR(*hpage)) {
 757                 if (!*wait)
 758                         return false;
 759
 760                 *wait = false;
 761                 *hpage = NULL;
 762                 khugepaged_alloc_sleep();
 763         } else if (*hpage) {
 764                 put_page(*hpage);
 765                 *hpage = NULL;
 766         }
 767
 768         return true;
 769 }
 770
 771 static struct page *
 772 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 773 {
 774         VM_BUG_ON_PAGE(*hpage, *hpage);
 775
 776         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
 777         if (unlikely(!*hpage)) {
 778                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 779                 *hpage = ERR_PTR(-ENOMEM);
 780                 return NULL;
 781         }
 782
 783         prep_transhuge_page(*hpage);
 784         count_vm_event(THP_COLLAPSE_ALLOC);
 785         return *hpage;
 786 }
 787 #else
 788 static int khugepaged_find_target_node(void)
 789 {
 790         return 0;
 791 }
 792
 793 static inline struct page *alloc_khugepaged_hugepage(void)
 794 {
 795         struct page *page;
 796
 797         page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
 798                            HPAGE_PMD_ORDER);
 799         if (page)
 800                 prep_transhuge_page(page);
 801         return page;
 802 }
 803
 804 static struct page *khugepaged_alloc_hugepage(bool *wait)
 805 {
 806         struct page *hpage;
 807
 808         do {
 809                 hpage = alloc_khugepaged_hugepage();
 810                 if (!hpage) {
 811                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 812                         if (!*wait)
 813                                 return NULL;
 814
 815                         *wait = false;
 816                         khugepaged_alloc_sleep();
 817                 } else
 818                         count_vm_event(THP_COLLAPSE_ALLOC);
 819         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
 820
 821         return hpage;
 822 }
 823
 824 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 825 {
 826         /*
 827          * If the hpage allocated earlier was briefly exposed in page cache
 828          * before collapse_file() failed, it is possible that racing lookups
 829          * have not yet completed, and would then be unpleasantly surprised by
 830          * finding the hpage reused for the same mapping at a different offset.
 831          * Just release the previous allocation if there is any danger of that.
 832          */
 833         if (*hpage && page_count(*hpage) > 1) {
 834                 put_page(*hpage);
 835                 *hpage = NULL;
 836         }
 837
 838         if (!*hpage)
 839                 *hpage = khugepaged_alloc_hugepage(wait);
 840
 841         if (unlikely(!*hpage))
 842                 return false;
 843
 844         return true;
 845 }
 846
 847 static struct page *
 848 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 849 {
 850         VM_BUG_ON(!*hpage);
 851
 852         return  *hpage;
 853 }
 854 #endif
 855
 856 /*
 857  * If mmap_sem temporarily dropped, revalidate vma
 858  * before taking mmap_sem.
 859  * Return 0 if succeeds, otherwise return none-zero
 860  * value (scan code).
 861  */
 862
 863 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
 864                 struct vm_area_struct **vmap)
 865 {
 866         struct vm_area_struct *vma;
 867         unsigned long hstart, hend;
 868
 869         if (unlikely(khugepaged_test_exit(mm)))
 870                 return SCAN_ANY_PROCESS;
 871
 872         *vmap = vma = find_vma(mm, address);
 873         if (!vma)
 874                 return SCAN_VMA_NULL;
 875
 876         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 877         hend = vma->vm_end & HPAGE_PMD_MASK;
 878         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
 879                 return SCAN_ADDRESS_RANGE;
 880         if (!hugepage_vma_check(vma, vma->vm_flags))
 881                 return SCAN_VMA_CHECK;
 882         return 0;
 883 }
 884
 885 /*
 886  * Bring missing pages in from swap, to complete THP collapse.
 887  * Only done if khugepaged_scan_pmd believes it is worthwhile.
 888  *
 889  * Called and returns without pte mapped or spinlocks held,
 890  * but with mmap_sem held to protect against vma changes.
 891  */
 892
 893 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
 894                                         struct vm_area_struct *vma,
 895                                         unsigned long address, pmd_t *pmd,
 896                                         int referenced)
 897 {
 898         int swapped_in = 0;
 899         vm_fault_t ret = 0;
 900         struct vm_fault vmf = {
 901                 .vma = vma,
 902                 .address = address,
 903                 .flags = FAULT_FLAG_ALLOW_RETRY,
 904                 .pmd = pmd,
 905                 .pgoff = linear_page_index(vma, address),
 906         };
 907
 908         /* we only decide to swapin, if there is enough young ptes */
 909         if (referenced < HPAGE_PMD_NR/2) {
 910                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 911                 return false;
 912         }
 913         vmf.pte = pte_offset_map(pmd, address);
 914         for (; vmf.address < address + HPAGE_PMD_NR*PAGE_SIZE;
 915                         vmf.pte++, vmf.address += PAGE_SIZE) {
 916                 vmf.orig_pte = *vmf.pte;
 917                 if (!is_swap_pte(vmf.orig_pte))
 918                         continue;
 919                 swapped_in++;
 920                 ret = do_swap_page(&vmf);
 921
 922                 /* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */
 923                 if (ret & VM_FAULT_RETRY) {
 924                         down_read(&mm->mmap_sem);
 925                         if (hugepage_vma_revalidate(mm, address, &vmf.vma)) {
 926                                 /* vma is no longer available, don't continue to swapin */
 927                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 928                                 return false;
 929                         }
 930                         /* check if the pmd is still valid */
 931                         if (mm_find_pmd(mm, address) != pmd) {
 932                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 933                                 return false;
 934                         }
 935                 }
 936                 if (ret & VM_FAULT_ERROR) {
 937                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 938                         return false;
 939                 }
 940                 /* pte is unmapped now, we need to map it */
 941                 vmf.pte = pte_offset_map(pmd, vmf.address);
 942         }
 943         vmf.pte--;
 944         pte_unmap(vmf.pte);
 945         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
 946         return true;
 947 }
 948
 949 static void collapse_huge_page(struct mm_struct *mm,
 950                                    unsigned long address,
 951                                    struct page **hpage,
 952                                    int node, int referenced)
 953 {
 954         pmd_t *pmd, _pmd;
 955         pte_t *pte;
 956         pgtable_t pgtable;
 957         struct page *new_page;
 958         spinlock_t *pmd_ptl, *pte_ptl;
 959         int isolated = 0, result = 0;
 960         struct mem_cgroup *memcg;
 961         struct vm_area_struct *vma;
 962         unsigned long mmun_start;       /* For mmu_notifiers */
 963         unsigned long mmun_end;         /* For mmu_notifiers */
 964         gfp_t gfp;
 965
 966         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 967
 968         /* Only allocate from the target node */
 969         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
 970
 971         /*
 972          * Before allocating the hugepage, release the mmap_sem read lock.
 973          * The allocation can take potentially a long time if it involves
 974          * sync compaction, and we do not need to hold the mmap_sem during
 975          * that. We will recheck the vma after taking it again in write mode.
 976          */
 977         up_read(&mm->mmap_sem);
 978         new_page = khugepaged_alloc_page(hpage, gfp, node);
 979         if (!new_page) {
 980                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
 981                 goto out_nolock;
 982         }
 983
 984         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
 985                 result = SCAN_CGROUP_CHARGE_FAIL;
 986                 goto out_nolock;
 987         }
 988
 989         down_read(&mm->mmap_sem);
 990         result = hugepage_vma_revalidate(mm, address, &vma);
 991         if (result) {
 992                 mem_cgroup_cancel_charge(new_page, memcg, true);
 993                 up_read(&mm->mmap_sem);
 994                 goto out_nolock;
 995         }
 996
 997         pmd = mm_find_pmd(mm, address);
 998         if (!pmd) {
 999                 result = SCAN_PMD_NULL;
1000                 mem_cgroup_cancel_charge(new_page, memcg, true);
1001                 up_read(&mm->mmap_sem);
1002                 goto out_nolock;
1003         }
1004
1005         /*
1006          * __collapse_huge_page_swapin always returns with mmap_sem locked.
1007          * If it fails, we release mmap_sem and jump out_nolock.
1008          * Continuing to collapse causes inconsistency.
1009          */
1010         if (!__collapse_huge_page_swapin(mm, vma, address, pmd, referenced)) {
1011                 mem_cgroup_cancel_charge(new_page, memcg, true);
1012                 up_read(&mm->mmap_sem);
1013                 goto out_nolock;
1014         }
1015
1016         up_read(&mm->mmap_sem);
1017         /*
1018          * Prevent all access to pagetables with the exception of
1019          * gup_fast later handled by the ptep_clear_flush and the VM
1020          * handled by the anon_vma lock + PG_lock.
1021          */
1022         down_write(&mm->mmap_sem);
1023         result = hugepage_vma_revalidate(mm, address, &vma);
1024         if (result)
1025                 goto out;
1026         /* check if the pmd is still valid */
1027         if (mm_find_pmd(mm, address) != pmd)
1028                 goto out;
1029
1030         anon_vma_lock_write(vma->anon_vma);
1031
1032         pte = pte_offset_map(pmd, address);
1033         pte_ptl = pte_lockptr(mm, pmd);
1034
1035         mmun_start = address;
1036         mmun_end   = address + HPAGE_PMD_SIZE;
1037         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1038         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1039         /*
1040          * After this gup_fast can't run anymore. This also removes
1041          * any huge TLB entry from the CPU so we won't allow
1042          * huge and small TLB entries for the same virtual address
1043          * to avoid the risk of CPU bugs in that area.
1044          */
1045         _pmd = pmdp_collapse_flush(vma, address, pmd);
1046         spin_unlock(pmd_ptl);
1047         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1048
1049         spin_lock(pte_ptl);
1050         isolated = __collapse_huge_page_isolate(vma, address, pte);
1051         spin_unlock(pte_ptl);
1052
1053         if (unlikely(!isolated)) {
1054                 pte_unmap(pte);
1055                 spin_lock(pmd_ptl);
1056                 BUG_ON(!pmd_none(*pmd));
1057                 /*
1058                  * We can only use set_pmd_at when establishing
1059                  * hugepmds and never for establishing regular pmds that
1060                  * points to regular pagetables. Use pmd_populate for that
1061                  */
1062                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1063                 spin_unlock(pmd_ptl);
1064                 anon_vma_unlock_write(vma->anon_vma);
1065                 result = SCAN_FAIL;
1066                 goto out;
1067         }
1068
1069         /*
1070          * All pages are isolated and locked so anon_vma rmap
1071          * can't run anymore.
1072          */
1073         anon_vma_unlock_write(vma->anon_vma);
1074
1075         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
1076         pte_unmap(pte);
1077         __SetPageUptodate(new_page);
1078         pgtable = pmd_pgtable(_pmd);
1079
1080         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1081         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1082
1083         /*
1084          * spin_lock() below is not the equivalent of smp_wmb(), so
1085          * this is needed to avoid the copy_huge_page writes to become
1086          * visible after the set_pmd_at() write.
1087          */
1088         smp_wmb();
1089
1090         spin_lock(pmd_ptl);
1091         BUG_ON(!pmd_none(*pmd));
1092         page_add_new_anon_rmap(new_page, vma, address, true);
1093         mem_cgroup_commit_charge(new_page, memcg, false, true);
1094         lru_cache_add_active_or_unevictable(new_page, vma);
1095         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1096         set_pmd_at(mm, address, pmd, _pmd);
1097         update_mmu_cache_pmd(vma, address, pmd);
1098         spin_unlock(pmd_ptl);
1099
1100         *hpage = NULL;
1101
1102         khugepaged_pages_collapsed++;
1103         result = SCAN_SUCCEED;
1104 out_up_write:
1105         up_write(&mm->mmap_sem);
1106 out_nolock:
1107         trace_mm_collapse_huge_page(mm, isolated, result);
1108         return;
1109 out:
1110         mem_cgroup_cancel_charge(new_page, memcg, true);
1111         goto out_up_write;
1112 }
1113
1114 static int khugepaged_scan_pmd(struct mm_struct *mm,
1115                                struct vm_area_struct *vma,
1116                                unsigned long address,
1117                                struct page **hpage)
1118 {
1119         pmd_t *pmd;
1120         pte_t *pte, *_pte;
1121         int ret = 0, none_or_zero = 0, result = 0, referenced = 0;
1122         struct page *page = NULL;
1123         unsigned long _address;
1124         spinlock_t *ptl;
1125         int node = NUMA_NO_NODE, unmapped = 0;
1126         bool writable = false;
1127
1128         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1129
1130         pmd = mm_find_pmd(mm, address);
1131         if (!pmd) {
1132                 result = SCAN_PMD_NULL;
1133                 goto out;
1134         }
1135
1136         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1137         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1138         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1139              _pte++, _address += PAGE_SIZE) {
1140                 pte_t pteval = *_pte;
1141                 if (is_swap_pte(pteval)) {
1142                         if (++unmapped <= khugepaged_max_ptes_swap) {
1143                                 continue;
1144                         } else {
1145                                 result = SCAN_EXCEED_SWAP_PTE;
1146                                 goto out_unmap;
1147                         }
1148                 }
1149                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1150                         if (!userfaultfd_armed(vma) &&
1151                             ++none_or_zero <= khugepaged_max_ptes_none) {
1152                                 continue;
1153                         } else {
1154                                 result = SCAN_EXCEED_NONE_PTE;
1155                                 goto out_unmap;
1156                         }
1157                 }
1158                 if (!pte_present(pteval)) {
1159                         result = SCAN_PTE_NON_PRESENT;
1160                         goto out_unmap;
1161                 }
1162                 if (pte_write(pteval))
1163                         writable = true;
1164
1165                 page = vm_normal_page(vma, _address, pteval);
1166                 if (unlikely(!page)) {
1167                         result = SCAN_PAGE_NULL;
1168                         goto out_unmap;
1169                 }
1170
1171                 /* TODO: teach khugepaged to collapse THP mapped with pte */
1172                 if (PageCompound(page)) {
1173                         result = SCAN_PAGE_COMPOUND;
1174                         goto out_unmap;
1175                 }
1176
1177                 /*
1178                  * Record which node the original page is from and save this
1179                  * information to khugepaged_node_load[].
1180                  * Khupaged will allocate hugepage from the node has the max
1181                  * hit record.
1182                  */
1183                 node = page_to_nid(page);
1184                 if (khugepaged_scan_abort(node)) {
1185                         result = SCAN_SCAN_ABORT;
1186                         goto out_unmap;
1187                 }
1188                 khugepaged_node_load[node]++;
1189                 if (!PageLRU(page)) {
1190                         result = SCAN_PAGE_LRU;
1191                         goto out_unmap;
1192                 }
1193                 if (PageLocked(page)) {
1194                         result = SCAN_PAGE_LOCK;
1195                         goto out_unmap;
1196                 }
1197                 if (!PageAnon(page)) {
1198                         result = SCAN_PAGE_ANON;
1199                         goto out_unmap;
1200                 }
1201
1202                 /*
1203                  * cannot use mapcount: can't collapse if there's a gup pin.
1204                  * The page must only be referenced by the scanned process
1205                  * and page swap cache.
1206                  */
1207                 if (page_count(page) != 1 + PageSwapCache(page)) {
1208                         result = SCAN_PAGE_COUNT;
1209                         goto out_unmap;
1210                 }
1211                 if (pte_young(pteval) ||
1212                     page_is_young(page) || PageReferenced(page) ||
1213                     mmu_notifier_test_young(vma->vm_mm, address))
1214                         referenced++;
1215         }
1216         if (writable) {
1217                 if (referenced) {
1218                         result = SCAN_SUCCEED;
1219                         ret = 1;
1220                 } else {
1221                         result = SCAN_LACK_REFERENCED_PAGE;
1222                 }
1223         } else {
1224                 result = SCAN_PAGE_RO;
1225         }
1226 out_unmap:
1227         pte_unmap_unlock(pte, ptl);
1228         if (ret) {
1229                 node = khugepaged_find_target_node();
1230                 /* collapse_huge_page will return with the mmap_sem released */
1231                 collapse_huge_page(mm, address, hpage, node, referenced);
1232         }
1233 out:
1234         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1235                                      none_or_zero, result, unmapped);
1236         return ret;
1237 }
1238
1239 static void collect_mm_slot(struct mm_slot *mm_slot)
1240 {
1241         struct mm_struct *mm = mm_slot->mm;
1242
1243         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
1244
1245         if (khugepaged_test_exit(mm)) {
1246                 /* free mm_slot */
1247                 hash_del(&mm_slot->hash);
1248                 list_del(&mm_slot->mm_node);
1249
1250                 /*
1251                  * Not strictly needed because the mm exited already.
1252                  *
1253                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1254                  */
1255
1256                 /* khugepaged_mm_lock actually not necessary for the below */
1257                 free_mm_slot(mm_slot);
1258                 mmdrop(mm);
1259         }
1260 }
1261
1262 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
1263 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1264 {
1265         struct vm_area_struct *vma;
1266         struct mm_struct *mm;
1267         unsigned long addr;
1268         pmd_t *pmd, _pmd;
1269
1270         i_mmap_lock_write(mapping);
1271         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1272                 /* probably overkill */
1273                 if (vma->anon_vma)
1274                         continue;
1275                 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1276                 if (addr & ~HPAGE_PMD_MASK)
1277                         continue;
1278                 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1279                         continue;
1280                 mm = vma->vm_mm;
1281                 pmd = mm_find_pmd(mm, addr);
1282                 if (!pmd)
1283                         continue;
1284                 /*
1285                  * We need exclusive mmap_sem to retract page table.
1286                  * If trylock fails we would end up with pte-mapped THP after
1287                  * re-fault. Not ideal, but it's more important to not disturb
1288                  * the system too much.
1289                  */
1290                 if (down_write_trylock(&mm->mmap_sem)) {
1291                         if (!khugepaged_test_exit(mm)) {
1292                                 spinlock_t *ptl = pmd_lock(mm, pmd);
1293                                 /* assume page table is clear */
1294                                 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1295                                 spin_unlock(ptl);
1296                                 mm_dec_nr_ptes(mm);
1297                                 pte_free(mm, pmd_pgtable(_pmd));
1298                         }
1299                         up_write(&mm->mmap_sem);
1300                 }
1301         }
1302         i_mmap_unlock_write(mapping);
1303 }
1304
1305 /**
1306  * collapse_shmem - collapse small tmpfs/shmem pages into huge one.
1307  *
1308  * Basic scheme is simple, details are more complex:
1309  *  - allocate and lock a new huge page;
1310  *  - scan over radix tree replacing old pages the new one
1311  *    + swap in pages if necessary;
1312  *    + fill in gaps;
1313  *    + keep old pages around in case if rollback is required;
1314  *  - if replacing succeed:
1315  *    + copy data over;
1316  *    + free old pages;
1317  *    + unlock huge page;
1318  *  - if replacing failed;
1319  *    + put all pages back and unfreeze them;
1320  *    + restore gaps in the radix-tree;
1321  *    + unlock and free huge page;
1322  */
1323 static void collapse_shmem(struct mm_struct *mm,
1324                 struct address_space *mapping, pgoff_t start,
1325                 struct page **hpage, int node)
1326 {
1327         gfp_t gfp;
1328         struct page *page, *new_page, *tmp;
1329         struct mem_cgroup *memcg;
1330         pgoff_t index, end = start + HPAGE_PMD_NR;
1331         LIST_HEAD(pagelist);
1332         struct radix_tree_iter iter;
1333         void **slot;
1334         int nr_none = 0, result = SCAN_SUCCEED;
1335
1336         VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1337
1338         /* Only allocate from the target node */
1339         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1340
1341         new_page = khugepaged_alloc_page(hpage, gfp, node);
1342         if (!new_page) {
1343                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1344                 goto out;
1345         }
1346
1347         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
1348                 result = SCAN_CGROUP_CHARGE_FAIL;
1349                 goto out;
1350         }
1351
1352         __SetPageLocked(new_page);
1353         __SetPageSwapBacked(new_page);
1354         new_page->index = start;
1355         new_page->mapping = mapping;
1356
1357         /*
1358          * At this point the new_page is locked and not up-to-date.
1359          * It's safe to insert it into the page cache, because nobody would
1360          * be able to map it or use it in another way until we unlock it.
1361          */
1362
1363         index = start;
1364         xa_lock_irq(&mapping->i_pages);
1365         radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start) {
1366                 int n = min(iter.index, end) - index;
1367
1368                 /*
1369                  * Stop if extent has been hole-punched, and is now completely
1370                  * empty (the more obvious i_size_read() check would take an
1371                  * irq-unsafe seqlock on 32-bit).
1372                  */
1373                 if (n >= HPAGE_PMD_NR) {
1374                         result = SCAN_TRUNCATED;
1375                         goto tree_locked;
1376                 }
1377
1378                 /*
1379                  * Handle holes in the radix tree: charge it from shmem and
1380                  * insert relevant subpage of new_page into the radix-tree.
1381                  */
1382                 if (n && !shmem_charge(mapping->host, n)) {
1383                         result = SCAN_FAIL;
1384                         goto tree_locked;
1385                 }
1386                 for (; index < min(iter.index, end); index++) {
1387                         radix_tree_insert(&mapping->i_pages, index,
1388                                         new_page + (index % HPAGE_PMD_NR));
1389                 }
1390                 nr_none += n;
1391
1392                 /* We are done. */
1393                 if (index >= end)
1394                         break;
1395
1396                 page = radix_tree_deref_slot_protected(slot,
1397                                 &mapping->i_pages.xa_lock);
1398                 if (radix_tree_exceptional_entry(page) || !PageUptodate(page)) {
1399                         xa_unlock_irq(&mapping->i_pages);
1400                         /* swap in or instantiate fallocated page */
1401                         if (shmem_getpage(mapping->host, index, &page,
1402                                                 SGP_NOHUGE)) {
1403                                 result = SCAN_FAIL;
1404                                 goto tree_unlocked;
1405                         }
1406                 } else if (trylock_page(page)) {
1407                         get_page(page);
1408                         xa_unlock_irq(&mapping->i_pages);
1409                 } else {
1410                         result = SCAN_PAGE_LOCK;
1411                         goto tree_locked;
1412                 }
1413
1414                 /*
1415                  * The page must be locked, so we can drop the i_pages lock
1416                  * without racing with truncate.
1417                  */
1418                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1419                 VM_BUG_ON_PAGE(!PageUptodate(page), page);
1420
1421                 /*
1422                  * If file was truncated then extended, or hole-punched, before
1423                  * we locked the first page, then a THP might be there already.
1424                  */
1425                 if (PageTransCompound(page)) {
1426                         result = SCAN_PAGE_COMPOUND;
1427                         goto out_unlock;
1428                 }
1429
1430                 if (page_mapping(page) != mapping) {
1431                         result = SCAN_TRUNCATED;
1432                         goto out_unlock;
1433                 }
1434
1435                 if (isolate_lru_page(page)) {
1436                         result = SCAN_DEL_PAGE_LRU;
1437                         goto out_unlock;
1438                 }
1439
1440                 if (page_mapped(page))
1441                         unmap_mapping_pages(mapping, index, 1, false);
1442
1443                 xa_lock_irq(&mapping->i_pages);
1444
1445                 slot = radix_tree_lookup_slot(&mapping->i_pages, index);
1446                 VM_BUG_ON_PAGE(page != radix_tree_deref_slot_protected(slot,
1447                                         &mapping->i_pages.xa_lock), page);
1448                 VM_BUG_ON_PAGE(page_mapped(page), page);
1449
1450                 /*
1451                  * The page is expected to have page_count() == 3:
1452                  *  - we hold a pin on it;
1453                  *  - one reference from radix tree;
1454                  *  - one from isolate_lru_page;
1455                  */
1456                 if (!page_ref_freeze(page, 3)) {
1457                         result = SCAN_PAGE_COUNT;
1458                         xa_unlock_irq(&mapping->i_pages);
1459                         putback_lru_page(page);
1460                         goto out_unlock;
1461                 }
1462
1463                 /*
1464                  * Add the page to the list to be able to undo the collapse if
1465                  * something go wrong.
1466                  */
1467                 list_add_tail(&page->lru, &pagelist);
1468
1469                 /* Finally, replace with the new page. */
1470                 radix_tree_replace_slot(&mapping->i_pages, slot,
1471                                 new_page + (index % HPAGE_PMD_NR));
1472
1473                 slot = radix_tree_iter_resume(slot, &iter);
1474                 index++;
1475                 continue;
1476 out_unlock:
1477                 unlock_page(page);
1478                 put_page(page);
1479                 goto tree_unlocked;
1480         }
1481
1482         /*
1483          * Handle hole in radix tree at the end of the range.
1484          * This code only triggers if there's nothing in radix tree
1485          * beyond 'end'.
1486          */
1487         if (index < end) {
1488                 int n = end - index;
1489
1490                 /* Stop if extent has been truncated, and is now empty */
1491                 if (n >= HPAGE_PMD_NR) {
1492                         result = SCAN_TRUNCATED;
1493                         goto tree_locked;
1494                 }
1495                 if (!shmem_charge(mapping->host, n)) {
1496                         result = SCAN_FAIL;
1497                         goto tree_locked;
1498                 }
1499                 for (; index < end; index++) {
1500                         radix_tree_insert(&mapping->i_pages, index,
1501                                         new_page + (index % HPAGE_PMD_NR));
1502                 }
1503                 nr_none += n;
1504         }
1505
1506         __inc_node_page_state(new_page, NR_SHMEM_THPS);
1507         if (nr_none) {
1508                 struct zone *zone = page_zone(new_page);
1509
1510                 __mod_node_page_state(zone->zone_pgdat, NR_FILE_PAGES, nr_none);
1511                 __mod_node_page_state(zone->zone_pgdat, NR_SHMEM, nr_none);
1512         }
1513
1514 tree_locked:
1515         xa_unlock_irq(&mapping->i_pages);
1516 tree_unlocked:
1517
1518         if (result == SCAN_SUCCEED) {
1519                 /*
1520                  * Replacing old pages with new one has succeed, now we need to
1521                  * copy the content and free old pages.
1522                  */
1523                 index = start;
1524                 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1525                         while (index < page->index) {
1526                                 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1527                                 index++;
1528                         }
1529                         copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1530                                         page);
1531                         list_del(&page->lru);
1532                         page->mapping = NULL;
1533                         page_ref_unfreeze(page, 1);
1534                         ClearPageActive(page);
1535                         ClearPageUnevictable(page);
1536                         unlock_page(page);
1537                         put_page(page);
1538                         index++;
1539                 }
1540                 while (index < end) {
1541                         clear_highpage(new_page + (index % HPAGE_PMD_NR));
1542                         index++;
1543                 }
1544
1545                 SetPageUptodate(new_page);
1546                 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1547                 set_page_dirty(new_page);
1548                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1549                 lru_cache_add_anon(new_page);
1550
1551                 /*
1552                  * Remove pte page tables, so we can re-fault the page as huge.
1553                  */
1554                 retract_page_tables(mapping, start);
1555                 *hpage = NULL;
1556
1557                 khugepaged_pages_collapsed++;
1558         } else {
1559                 /* Something went wrong: rollback changes to the radix-tree */
1560                 xa_lock_irq(&mapping->i_pages);
1561                 mapping->nrpages -= nr_none;
1562                 shmem_uncharge(mapping->host, nr_none);
1563
1564                 radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start) {
1565                         if (iter.index >= end)
1566                                 break;
1567                         page = list_first_entry_or_null(&pagelist,
1568                                         struct page, lru);
1569                         if (!page || iter.index < page->index) {
1570                                 if (!nr_none)
1571                                         break;
1572                                 nr_none--;
1573                                 /* Put holes back where they were */
1574                                 radix_tree_delete(&mapping->i_pages, iter.index);
1575                                 continue;
1576                         }
1577
1578                         VM_BUG_ON_PAGE(page->index != iter.index, page);
1579
1580                         /* Unfreeze the page. */
1581                         list_del(&page->lru);
1582                         page_ref_unfreeze(page, 2);
1583                         radix_tree_replace_slot(&mapping->i_pages, slot, page);
1584                         slot = radix_tree_iter_resume(slot, &iter);
1585                         xa_unlock_irq(&mapping->i_pages);
1586                         unlock_page(page);
1587                         putback_lru_page(page);
1588                         xa_lock_irq(&mapping->i_pages);
1589                 }
1590                 VM_BUG_ON(nr_none);
1591                 xa_unlock_irq(&mapping->i_pages);
1592
1593                 mem_cgroup_cancel_charge(new_page, memcg, true);
1594                 new_page->mapping = NULL;
1595         }
1596
1597         unlock_page(new_page);
1598 out:
1599         VM_BUG_ON(!list_empty(&pagelist));
1600         /* TODO: tracepoints */
1601 }
1602
1603 static void khugepaged_scan_shmem(struct mm_struct *mm,
1604                 struct address_space *mapping,
1605                 pgoff_t start, struct page **hpage)
1606 {
1607         struct page *page = NULL;
1608         struct radix_tree_iter iter;
1609         void **slot;
1610         int present, swap;
1611         int node = NUMA_NO_NODE;
1612         int result = SCAN_SUCCEED;
1613
1614         present = 0;
1615         swap = 0;
1616         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1617         rcu_read_lock();
1618         radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start) {
1619                 if (iter.index >= start + HPAGE_PMD_NR)
1620                         break;
1621
1622                 page = radix_tree_deref_slot(slot);
1623                 if (radix_tree_deref_retry(page)) {
1624                         slot = radix_tree_iter_retry(&iter);
1625                         continue;
1626                 }
1627
1628                 if (radix_tree_exception(page)) {
1629                         if (++swap > khugepaged_max_ptes_swap) {
1630                                 result = SCAN_EXCEED_SWAP_PTE;
1631                                 break;
1632                         }
1633                         continue;
1634                 }
1635
1636                 if (PageTransCompound(page)) {
1637                         result = SCAN_PAGE_COMPOUND;
1638                         break;
1639                 }
1640
1641                 node = page_to_nid(page);
1642                 if (khugepaged_scan_abort(node)) {
1643                         result = SCAN_SCAN_ABORT;
1644                         break;
1645                 }
1646                 khugepaged_node_load[node]++;
1647
1648                 if (!PageLRU(page)) {
1649                         result = SCAN_PAGE_LRU;
1650                         break;
1651                 }
1652
1653                 if (page_count(page) != 1 + page_mapcount(page)) {
1654                         result = SCAN_PAGE_COUNT;
1655                         break;
1656                 }
1657
1658                 /*
1659                  * We probably should check if the page is referenced here, but
1660                  * nobody would transfer pte_young() to PageReferenced() for us.
1661                  * And rmap walk here is just too costly...
1662                  */
1663
1664                 present++;
1665
1666                 if (need_resched()) {
1667                         slot = radix_tree_iter_resume(slot, &iter);
1668                         cond_resched_rcu();
1669                 }
1670         }
1671         rcu_read_unlock();
1672
1673         if (result == SCAN_SUCCEED) {
1674                 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
1675                         result = SCAN_EXCEED_NONE_PTE;
1676                 } else {
1677                         node = khugepaged_find_target_node();
1678                         collapse_shmem(mm, mapping, start, hpage, node);
1679                 }
1680         }
1681
1682         /* TODO: tracepoints */
1683 }
1684 #else
1685 static void khugepaged_scan_shmem(struct mm_struct *mm,
1686                 struct address_space *mapping,
1687                 pgoff_t start, struct page **hpage)
1688 {
1689         BUILD_BUG();
1690 }
1691 #endif
1692
1693 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
1694                                             struct page **hpage)
1695         __releases(&khugepaged_mm_lock)
1696         __acquires(&khugepaged_mm_lock)
1697 {
1698         struct mm_slot *mm_slot;
1699         struct mm_struct *mm;
1700         struct vm_area_struct *vma;
1701         int progress = 0;
1702
1703         VM_BUG_ON(!pages);
1704         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
1705
1706         if (khugepaged_scan.mm_slot)
1707                 mm_slot = khugepaged_scan.mm_slot;
1708         else {
1709                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
1710                                      struct mm_slot, mm_node);
1711                 khugepaged_scan.address = 0;
1712                 khugepaged_scan.mm_slot = mm_slot;
1713         }
1714         spin_unlock(&khugepaged_mm_lock);
1715
1716         mm = mm_slot->mm;
1717         /*
1718          * Don't wait for semaphore (to avoid long wait times).  Just move to
1719          * the next mm on the list.
1720          */
1721         vma = NULL;
1722         if (unlikely(!down_read_trylock(&mm->mmap_sem)))
1723                 goto breakouterloop_mmap_sem;
1724         if (likely(!khugepaged_test_exit(mm)))
1725                 vma = find_vma(mm, khugepaged_scan.address);
1726
1727         progress++;
1728         for (; vma; vma = vma->vm_next) {
1729                 unsigned long hstart, hend;
1730
1731                 cond_resched();
1732                 if (unlikely(khugepaged_test_exit(mm))) {
1733                         progress++;
1734                         break;
1735                 }
1736                 if (!hugepage_vma_check(vma, vma->vm_flags)) {
1737 skip:
1738                         progress++;
1739                         continue;
1740                 }
1741                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1742                 hend = vma->vm_end & HPAGE_PMD_MASK;
1743                 if (hstart >= hend)
1744                         goto skip;
1745                 if (khugepaged_scan.address > hend)
1746                         goto skip;
1747                 if (khugepaged_scan.address < hstart)
1748                         khugepaged_scan.address = hstart;
1749                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
1750
1751                 while (khugepaged_scan.address < hend) {
1752                         int ret;
1753                         cond_resched();
1754                         if (unlikely(khugepaged_test_exit(mm)))
1755                                 goto breakouterloop;
1756
1757                         VM_BUG_ON(khugepaged_scan.address < hstart ||
1758                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
1759                                   hend);
1760                         if (shmem_file(vma->vm_file)) {
1761                                 struct file *file;
1762                                 pgoff_t pgoff = linear_page_index(vma,
1763                                                 khugepaged_scan.address);
1764                                 if (!shmem_huge_enabled(vma))
1765                                         goto skip;
1766                                 file = get_file(vma->vm_file);
1767                                 up_read(&mm->mmap_sem);
1768                                 ret = 1;
1769                                 khugepaged_scan_shmem(mm, file->f_mapping,
1770                                                 pgoff, hpage);
1771                                 fput(file);
1772                         } else {
1773                                 ret = khugepaged_scan_pmd(mm, vma,
1774                                                 khugepaged_scan.address,
1775                                                 hpage);
1776                         }
1777                         /* move to next address */
1778                         khugepaged_scan.address += HPAGE_PMD_SIZE;
1779                         progress += HPAGE_PMD_NR;
1780                         if (ret)
1781                                 /* we released mmap_sem so break loop */
1782                                 goto breakouterloop_mmap_sem;
1783                         if (progress >= pages)
1784                                 goto breakouterloop;
1785                 }
1786         }
1787 breakouterloop:
1788         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
1789 breakouterloop_mmap_sem:
1790
1791         spin_lock(&khugepaged_mm_lock);
1792         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
1793         /*
1794          * Release the current mm_slot if this mm is about to die, or
1795          * if we scanned all vmas of this mm.
1796          */
1797         if (khugepaged_test_exit(mm) || !vma) {
1798                 /*
1799                  * Make sure that if mm_users is reaching zero while
1800                  * khugepaged runs here, khugepaged_exit will find
1801                  * mm_slot not pointing to the exiting mm.
1802                  */
1803                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
1804                         khugepaged_scan.mm_slot = list_entry(
1805                                 mm_slot->mm_node.next,
1806                                 struct mm_slot, mm_node);
1807                         khugepaged_scan.address = 0;
1808                 } else {
1809                         khugepaged_scan.mm_slot = NULL;
1810                         khugepaged_full_scans++;
1811                 }
1812
1813                 collect_mm_slot(mm_slot);
1814         }
1815
1816         return progress;
1817 }
1818
1819 static int khugepaged_has_work(void)
1820 {
1821         return !list_empty(&khugepaged_scan.mm_head) &&
1822                 khugepaged_enabled();
1823 }
1824
1825 static int khugepaged_wait_event(void)
1826 {
1827         return !list_empty(&khugepaged_scan.mm_head) ||
1828                 kthread_should_stop();
1829 }
1830
1831 static void khugepaged_do_scan(void)
1832 {
1833         struct page *hpage = NULL;
1834         unsigned int progress = 0, pass_through_head = 0;
1835         unsigned int pages = khugepaged_pages_to_scan;
1836         bool wait = true;
1837
1838         barrier(); /* write khugepaged_pages_to_scan to local stack */
1839
1840         while (progress < pages) {
1841                 if (!khugepaged_prealloc_page(&hpage, &wait))
1842                         break;
1843
1844                 cond_resched();
1845
1846                 if (unlikely(kthread_should_stop() || try_to_freeze()))
1847                         break;
1848
1849                 spin_lock(&khugepaged_mm_lock);
1850                 if (!khugepaged_scan.mm_slot)
1851                         pass_through_head++;
1852                 if (khugepaged_has_work() &&
1853                     pass_through_head < 2)
1854                         progress += khugepaged_scan_mm_slot(pages - progress,
1855                                                             &hpage);
1856                 else
1857                         progress = pages;
1858                 spin_unlock(&khugepaged_mm_lock);
1859         }
1860
1861         if (!IS_ERR_OR_NULL(hpage))
1862                 put_page(hpage);
1863 }
1864
1865 static bool khugepaged_should_wakeup(void)
1866 {
1867         return kthread_should_stop() ||
1868                time_after_eq(jiffies, khugepaged_sleep_expire);
1869 }
1870
1871 static void khugepaged_wait_work(void)
1872 {
1873         if (khugepaged_has_work()) {
1874                 const unsigned long scan_sleep_jiffies =
1875                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
1876
1877                 if (!scan_sleep_jiffies)
1878                         return;
1879
1880                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
1881                 wait_event_freezable_timeout(khugepaged_wait,
1882                                              khugepaged_should_wakeup(),
1883                                              scan_sleep_jiffies);
1884                 return;
1885         }
1886
1887         if (khugepaged_enabled())
1888                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
1889 }
1890
1891 static int khugepaged(void *none)
1892 {
1893         struct mm_slot *mm_slot;
1894
1895         set_freezable();
1896         set_user_nice(current, MAX_NICE);
1897
1898         while (!kthread_should_stop()) {
1899                 khugepaged_do_scan();
1900                 khugepaged_wait_work();
1901         }
1902
1903         spin_lock(&khugepaged_mm_lock);
1904         mm_slot = khugepaged_scan.mm_slot;
1905         khugepaged_scan.mm_slot = NULL;
1906         if (mm_slot)
1907                 collect_mm_slot(mm_slot);
1908         spin_unlock(&khugepaged_mm_lock);
1909         return 0;
1910 }
1911
1912 static void set_recommended_min_free_kbytes(void)
1913 {
1914         struct zone *zone;
1915         int nr_zones = 0;
1916         unsigned long recommended_min;
1917
1918         for_each_populated_zone(zone) {
1919                 /*
1920                  * We don't need to worry about fragmentation of
1921                  * ZONE_MOVABLE since it only has movable pages.
1922                  */
1923                 if (zone_idx(zone) > gfp_zone(GFP_USER))
1924                         continue;
1925
1926                 nr_zones++;
1927         }
1928
1929         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
1930         recommended_min = pageblock_nr_pages * nr_zones * 2;
1931
1932         /*
1933          * Make sure that on average at least two pageblocks are almost free
1934          * of another type, one for a migratetype to fall back to and a
1935          * second to avoid subsequent fallbacks of other types There are 3
1936          * MIGRATE_TYPES we care about.
1937          */
1938         recommended_min += pageblock_nr_pages * nr_zones *
1939                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
1940
1941         /* don't ever allow to reserve more than 5% of the lowmem */
1942         recommended_min = min(recommended_min,
1943                               (unsigned long) nr_free_buffer_pages() / 20);
1944         recommended_min <<= (PAGE_SHIFT-10);
1945
1946         if (recommended_min > min_free_kbytes) {
1947                 if (user_min_free_kbytes >= 0)
1948                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
1949                                 min_free_kbytes, recommended_min);
1950
1951                 min_free_kbytes = recommended_min;
1952         }
1953         setup_per_zone_wmarks();
1954 }
1955
1956 int start_stop_khugepaged(void)
1957 {
1958         int err = 0;
1959
1960         mutex_lock(&khugepaged_mutex);
1961         if (khugepaged_enabled()) {
1962                 if (!khugepaged_thread)
1963                         khugepaged_thread = kthread_run(khugepaged, NULL,
1964                                                         "khugepaged");
1965                 if (IS_ERR(khugepaged_thread)) {
1966                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
1967                         err = PTR_ERR(khugepaged_thread);
1968                         khugepaged_thread = NULL;
1969                         goto fail;
1970                 }
1971
1972                 if (!list_empty(&khugepaged_scan.mm_head))
1973                         wake_up_interruptible(&khugepaged_wait);
1974
1975                 set_recommended_min_free_kbytes();
1976         } else if (khugepaged_thread) {
1977                 kthread_stop(khugepaged_thread);
1978                 khugepaged_thread = NULL;
1979         }
1980 fail:
1981         mutex_unlock(&khugepaged_mutex);
1982         return err;
1983 }
1984
1985 void khugepaged_min_free_kbytes_update(void)
1986 {
1987         mutex_lock(&khugepaged_mutex);
1988         if (khugepaged_enabled() && khugepaged_thread)
1989                 set_recommended_min_free_kbytes();
1990         mutex_unlock(&khugepaged_mutex);
1991 }