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 /* gross hack for <=4.19 stable */
  27 #if defined(CONFIG_S390) || defined(CONFIG_ARM)
  28 static void tlb_remove_table_smp_sync(void *arg)
  29 {
  30         /* Simply deliver the interrupt */
  31 }
  32
  33 static void tlb_remove_table_sync_one(void)
  34 {
  35         smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
  36 }
  37 #endif
  38
  39 enum scan_result {
  40         SCAN_FAIL,
  41         SCAN_SUCCEED,
  42         SCAN_PMD_NULL,
  43         SCAN_EXCEED_NONE_PTE,
  44         SCAN_PTE_NON_PRESENT,
  45         SCAN_PAGE_RO,
  46         SCAN_LACK_REFERENCED_PAGE,
  47         SCAN_PAGE_NULL,
  48         SCAN_SCAN_ABORT,
  49         SCAN_PAGE_COUNT,
  50         SCAN_PAGE_LRU,
  51         SCAN_PAGE_LOCK,
  52         SCAN_PAGE_ANON,
  53         SCAN_PAGE_COMPOUND,
  54         SCAN_ANY_PROCESS,
  55         SCAN_VMA_NULL,
  56         SCAN_VMA_CHECK,
  57         SCAN_ADDRESS_RANGE,
  58         SCAN_SWAP_CACHE_PAGE,
  59         SCAN_DEL_PAGE_LRU,
  60         SCAN_ALLOC_HUGE_PAGE_FAIL,
  61         SCAN_CGROUP_CHARGE_FAIL,
  62         SCAN_EXCEED_SWAP_PTE,
  63         SCAN_TRUNCATED,
  64 };
  65
  66 #define CREATE_TRACE_POINTS
  67 #include <trace/events/huge_memory.h>
  68
  69 static struct task_struct *khugepaged_thread __read_mostly;
  70 static DEFINE_MUTEX(khugepaged_mutex);
  71
  72 /* default scan 8*512 pte (or vmas) every 30 second */
  73 static unsigned int khugepaged_pages_to_scan __read_mostly;
  74 static unsigned int khugepaged_pages_collapsed;
  75 static unsigned int khugepaged_full_scans;
  76 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  77 /* during fragmentation poll the hugepage allocator once every minute */
  78 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  79 static unsigned long khugepaged_sleep_expire;
  80 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  81 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  82 /*
  83  * default collapse hugepages if there is at least one pte mapped like
  84  * it would have happened if the vma was large enough during page
  85  * fault.
  86  */
  87 static unsigned int khugepaged_max_ptes_none __read_mostly;
  88 static unsigned int khugepaged_max_ptes_swap __read_mostly;
  89
  90 #define MM_SLOTS_HASH_BITS 10
  91 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  92
  93 static struct kmem_cache *mm_slot_cache __read_mostly;
  94
  95 /**
  96  * struct mm_slot - hash lookup from mm to mm_slot
  97  * @hash: hash collision list
  98  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  99  * @mm: the mm that this information is valid for
 100  */
 101 struct mm_slot {
 102         struct hlist_node hash;
 103         struct list_head mm_node;
 104         struct mm_struct *mm;
 105 };
 106
 107 /**
 108  * struct khugepaged_scan - cursor for scanning
 109  * @mm_head: the head of the mm list to scan
 110  * @mm_slot: the current mm_slot we are scanning
 111  * @address: the next address inside that to be scanned
 112  *
 113  * There is only the one khugepaged_scan instance of this cursor structure.
 114  */
 115 struct khugepaged_scan {
 116         struct list_head mm_head;
 117         struct mm_slot *mm_slot;
 118         unsigned long address;
 119 };
 120
 121 static struct khugepaged_scan khugepaged_scan = {
 122         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 123 };
 124
 125 #ifdef CONFIG_SYSFS
 126 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 127                                          struct kobj_attribute *attr,
 128                                          char *buf)
 129 {
 130         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 131 }
 132
 133 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 134                                           struct kobj_attribute *attr,
 135                                           const char *buf, size_t count)
 136 {
 137         unsigned long msecs;
 138         int err;
 139
 140         err = kstrtoul(buf, 10, &msecs);
 141         if (err || msecs > UINT_MAX)
 142                 return -EINVAL;
 143
 144         khugepaged_scan_sleep_millisecs = msecs;
 145         khugepaged_sleep_expire = 0;
 146         wake_up_interruptible(&khugepaged_wait);
 147
 148         return count;
 149 }
 150 static struct kobj_attribute scan_sleep_millisecs_attr =
 151         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 152                scan_sleep_millisecs_store);
 153
 154 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 155                                           struct kobj_attribute *attr,
 156                                           char *buf)
 157 {
 158         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 159 }
 160
 161 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 162                                            struct kobj_attribute *attr,
 163                                            const char *buf, size_t count)
 164 {
 165         unsigned long msecs;
 166         int err;
 167
 168         err = kstrtoul(buf, 10, &msecs);
 169         if (err || msecs > UINT_MAX)
 170                 return -EINVAL;
 171
 172         khugepaged_alloc_sleep_millisecs = msecs;
 173         khugepaged_sleep_expire = 0;
 174         wake_up_interruptible(&khugepaged_wait);
 175
 176         return count;
 177 }
 178 static struct kobj_attribute alloc_sleep_millisecs_attr =
 179         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 180                alloc_sleep_millisecs_store);
 181
 182 static ssize_t pages_to_scan_show(struct kobject *kobj,
 183                                   struct kobj_attribute *attr,
 184                                   char *buf)
 185 {
 186         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 187 }
 188 static ssize_t pages_to_scan_store(struct kobject *kobj,
 189                                    struct kobj_attribute *attr,
 190                                    const char *buf, size_t count)
 191 {
 192         int err;
 193         unsigned long pages;
 194
 195         err = kstrtoul(buf, 10, &pages);
 196         if (err || !pages || pages > UINT_MAX)
 197                 return -EINVAL;
 198
 199         khugepaged_pages_to_scan = pages;
 200
 201         return count;
 202 }
 203 static struct kobj_attribute pages_to_scan_attr =
 204         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 205                pages_to_scan_store);
 206
 207 static ssize_t pages_collapsed_show(struct kobject *kobj,
 208                                     struct kobj_attribute *attr,
 209                                     char *buf)
 210 {
 211         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 212 }
 213 static struct kobj_attribute pages_collapsed_attr =
 214         __ATTR_RO(pages_collapsed);
 215
 216 static ssize_t full_scans_show(struct kobject *kobj,
 217                                struct kobj_attribute *attr,
 218                                char *buf)
 219 {
 220         return sprintf(buf, "%u\n", khugepaged_full_scans);
 221 }
 222 static struct kobj_attribute full_scans_attr =
 223         __ATTR_RO(full_scans);
 224
 225 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 226                                       struct kobj_attribute *attr, char *buf)
 227 {
 228         return single_hugepage_flag_show(kobj, attr, buf,
 229                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 230 }
 231 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 232                                        struct kobj_attribute *attr,
 233                                        const char *buf, size_t count)
 234 {
 235         return single_hugepage_flag_store(kobj, attr, buf, count,
 236                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 237 }
 238 static struct kobj_attribute khugepaged_defrag_attr =
 239         __ATTR(defrag, 0644, khugepaged_defrag_show,
 240                khugepaged_defrag_store);
 241
 242 /*
 243  * max_ptes_none controls if khugepaged should collapse hugepages over
 244  * any unmapped ptes in turn potentially increasing the memory
 245  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 246  * reduce the available free memory in the system as it
 247  * runs. Increasing max_ptes_none will instead potentially reduce the
 248  * free memory in the system during the khugepaged scan.
 249  */
 250 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 251                                              struct kobj_attribute *attr,
 252                                              char *buf)
 253 {
 254         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 255 }
 256 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 257                                               struct kobj_attribute *attr,
 258                                               const char *buf, size_t count)
 259 {
 260         int err;
 261         unsigned long max_ptes_none;
 262
 263         err = kstrtoul(buf, 10, &max_ptes_none);
 264         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 265                 return -EINVAL;
 266
 267         khugepaged_max_ptes_none = max_ptes_none;
 268
 269         return count;
 270 }
 271 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 272         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 273                khugepaged_max_ptes_none_store);
 274
 275 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
 276                                              struct kobj_attribute *attr,
 277                                              char *buf)
 278 {
 279         return sprintf(buf, "%u\n", khugepaged_max_ptes_swap);
 280 }
 281
 282 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
 283                                               struct kobj_attribute *attr,
 284                                               const char *buf, size_t count)
 285 {
 286         int err;
 287         unsigned long max_ptes_swap;
 288
 289         err  = kstrtoul(buf, 10, &max_ptes_swap);
 290         if (err || max_ptes_swap > HPAGE_PMD_NR-1)
 291                 return -EINVAL;
 292
 293         khugepaged_max_ptes_swap = max_ptes_swap;
 294
 295         return count;
 296 }
 297
 298 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
 299         __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
 300                khugepaged_max_ptes_swap_store);
 301
 302 static struct attribute *khugepaged_attr[] = {
 303         &khugepaged_defrag_attr.attr,
 304         &khugepaged_max_ptes_none_attr.attr,
 305         &pages_to_scan_attr.attr,
 306         &pages_collapsed_attr.attr,
 307         &full_scans_attr.attr,
 308         &scan_sleep_millisecs_attr.attr,
 309         &alloc_sleep_millisecs_attr.attr,
 310         &khugepaged_max_ptes_swap_attr.attr,
 311         NULL,
 312 };
 313
 314 struct attribute_group khugepaged_attr_group = {
 315         .attrs = khugepaged_attr,
 316         .name = "khugepaged",
 317 };
 318 #endif /* CONFIG_SYSFS */
 319
 320 #define VM_NO_KHUGEPAGED (VM_SPECIAL | VM_HUGETLB)
 321
 322 int hugepage_madvise(struct vm_area_struct *vma,
 323                      unsigned long *vm_flags, int advice)
 324 {
 325         switch (advice) {
 326         case MADV_HUGEPAGE:
 327 #ifdef CONFIG_S390
 328                 /*
 329                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
 330                  * can't handle this properly after s390_enable_sie, so we simply
 331                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
 332                  */
 333                 if (mm_has_pgste(vma->vm_mm))
 334                         return 0;
 335 #endif
 336                 *vm_flags &= ~VM_NOHUGEPAGE;
 337                 *vm_flags |= VM_HUGEPAGE;
 338                 /*
 339                  * If the vma become good for khugepaged to scan,
 340                  * register it here without waiting a page fault that
 341                  * may not happen any time soon.
 342                  */
 343                 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
 344                                 khugepaged_enter_vma_merge(vma, *vm_flags))
 345                         return -ENOMEM;
 346                 break;
 347         case MADV_NOHUGEPAGE:
 348                 *vm_flags &= ~VM_HUGEPAGE;
 349                 *vm_flags |= VM_NOHUGEPAGE;
 350                 /*
 351                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
 352                  * this vma even if we leave the mm registered in khugepaged if
 353                  * it got registered before VM_NOHUGEPAGE was set.
 354                  */
 355                 break;
 356         }
 357
 358         return 0;
 359 }
 360
 361 int __init khugepaged_init(void)
 362 {
 363         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
 364                                           sizeof(struct mm_slot),
 365                                           __alignof__(struct mm_slot), 0, NULL);
 366         if (!mm_slot_cache)
 367                 return -ENOMEM;
 368
 369         khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
 370         khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
 371         khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
 372
 373         return 0;
 374 }
 375
 376 void __init khugepaged_destroy(void)
 377 {
 378         kmem_cache_destroy(mm_slot_cache);
 379 }
 380
 381 static inline struct mm_slot *alloc_mm_slot(void)
 382 {
 383         if (!mm_slot_cache)     /* initialization failed */
 384                 return NULL;
 385         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
 386 }
 387
 388 static inline void free_mm_slot(struct mm_slot *mm_slot)
 389 {
 390         kmem_cache_free(mm_slot_cache, mm_slot);
 391 }
 392
 393 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
 394 {
 395         struct mm_slot *mm_slot;
 396
 397         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
 398                 if (mm == mm_slot->mm)
 399                         return mm_slot;
 400
 401         return NULL;
 402 }
 403
 404 static void insert_to_mm_slots_hash(struct mm_struct *mm,
 405                                     struct mm_slot *mm_slot)
 406 {
 407         mm_slot->mm = mm;
 408         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
 409 }
 410
 411 static inline int khugepaged_test_exit(struct mm_struct *mm)
 412 {
 413         return atomic_read(&mm->mm_users) == 0 || !mmget_still_valid(mm);
 414 }
 415
 416 static bool hugepage_vma_check(struct vm_area_struct *vma,
 417                                unsigned long vm_flags)
 418 {
 419         if ((!(vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
 420             (vm_flags & VM_NOHUGEPAGE) ||
 421             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
 422                 return false;
 423         if (shmem_file(vma->vm_file)) {
 424                 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
 425                         return false;
 426                 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
 427                                 HPAGE_PMD_NR);
 428         }
 429         if (!vma->anon_vma || vma->vm_ops)
 430                 return false;
 431         if (is_vma_temporary_stack(vma))
 432                 return false;
 433         return !(vm_flags & VM_NO_KHUGEPAGED);
 434 }
 435
 436 int __khugepaged_enter(struct mm_struct *mm)
 437 {
 438         struct mm_slot *mm_slot;
 439         int wakeup;
 440
 441         mm_slot = alloc_mm_slot();
 442         if (!mm_slot)
 443                 return -ENOMEM;
 444
 445         /* __khugepaged_exit() must not run from under us */
 446         VM_BUG_ON_MM(atomic_read(&mm->mm_users) == 0, mm);
 447         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
 448                 free_mm_slot(mm_slot);
 449                 return 0;
 450         }
 451
 452         spin_lock(&khugepaged_mm_lock);
 453         insert_to_mm_slots_hash(mm, mm_slot);
 454         /*
 455          * Insert just behind the scanning cursor, to let the area settle
 456          * down a little.
 457          */
 458         wakeup = list_empty(&khugepaged_scan.mm_head);
 459         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
 460         spin_unlock(&khugepaged_mm_lock);
 461
 462         mmgrab(mm);
 463         if (wakeup)
 464                 wake_up_interruptible(&khugepaged_wait);
 465
 466         return 0;
 467 }
 468
 469 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
 470                                unsigned long vm_flags)
 471 {
 472         unsigned long hstart, hend;
 473
 474         /*
 475          * khugepaged does not yet work on non-shmem files or special
 476          * mappings. And file-private shmem THP is not supported.
 477          */
 478         if (!hugepage_vma_check(vma, vm_flags))
 479                 return 0;
 480
 481         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 482         hend = vma->vm_end & HPAGE_PMD_MASK;
 483         if (hstart < hend)
 484                 return khugepaged_enter(vma, vm_flags);
 485         return 0;
 486 }
 487
 488 void __khugepaged_exit(struct mm_struct *mm)
 489 {
 490         struct mm_slot *mm_slot;
 491         int free = 0;
 492
 493         spin_lock(&khugepaged_mm_lock);
 494         mm_slot = get_mm_slot(mm);
 495         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
 496                 hash_del(&mm_slot->hash);
 497                 list_del(&mm_slot->mm_node);
 498                 free = 1;
 499         }
 500         spin_unlock(&khugepaged_mm_lock);
 501
 502         if (free) {
 503                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
 504                 free_mm_slot(mm_slot);
 505                 mmdrop(mm);
 506         } else if (mm_slot) {
 507                 /*
 508                  * This is required to serialize against
 509                  * khugepaged_test_exit() (which is guaranteed to run
 510                  * under mmap sem read mode). Stop here (after we
 511                  * return all pagetables will be destroyed) until
 512                  * khugepaged has finished working on the pagetables
 513                  * under the mmap_sem.
 514                  */
 515                 down_write(&mm->mmap_sem);
 516                 up_write(&mm->mmap_sem);
 517         }
 518 }
 519
 520 static void release_pte_page(struct page *page)
 521 {
 522         dec_node_page_state(page, NR_ISOLATED_ANON + page_is_file_cache(page));
 523         unlock_page(page);
 524         putback_lru_page(page);
 525 }
 526
 527 static void release_pte_pages(pte_t *pte, pte_t *_pte)
 528 {
 529         while (--_pte >= pte) {
 530                 pte_t pteval = *_pte;
 531                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
 532                         release_pte_page(pte_page(pteval));
 533         }
 534 }
 535
 536 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
 537                                         unsigned long address,
 538                                         pte_t *pte)
 539 {
 540         struct page *page = NULL;
 541         pte_t *_pte;
 542         int none_or_zero = 0, result = 0, referenced = 0;
 543         bool writable = false;
 544
 545         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
 546              _pte++, address += PAGE_SIZE) {
 547                 pte_t pteval = *_pte;
 548                 if (pte_none(pteval) || (pte_present(pteval) &&
 549                                 is_zero_pfn(pte_pfn(pteval)))) {
 550                         if (!userfaultfd_armed(vma) &&
 551                             ++none_or_zero <= khugepaged_max_ptes_none) {
 552                                 continue;
 553                         } else {
 554                                 result = SCAN_EXCEED_NONE_PTE;
 555                                 goto out;
 556                         }
 557                 }
 558                 if (!pte_present(pteval)) {
 559                         result = SCAN_PTE_NON_PRESENT;
 560                         goto out;
 561                 }
 562                 page = vm_normal_page(vma, address, pteval);
 563                 if (unlikely(!page)) {
 564                         result = SCAN_PAGE_NULL;
 565                         goto out;
 566                 }
 567
 568                 /* TODO: teach khugepaged to collapse THP mapped with pte */
 569                 if (PageCompound(page)) {
 570                         result = SCAN_PAGE_COMPOUND;
 571                         goto out;
 572                 }
 573
 574                 VM_BUG_ON_PAGE(!PageAnon(page), page);
 575
 576                 /*
 577                  * We can do it before isolate_lru_page because the
 578                  * page can't be freed from under us. NOTE: PG_lock
 579                  * is needed to serialize against split_huge_page
 580                  * when invoked from the VM.
 581                  */
 582                 if (!trylock_page(page)) {
 583                         result = SCAN_PAGE_LOCK;
 584                         goto out;
 585                 }
 586
 587                 /*
 588                  * cannot use mapcount: can't collapse if there's a gup pin.
 589                  * The page must only be referenced by the scanned process
 590                  * and page swap cache.
 591                  */
 592                 if (page_count(page) != 1 + PageSwapCache(page)) {
 593                         unlock_page(page);
 594                         result = SCAN_PAGE_COUNT;
 595                         goto out;
 596                 }
 597                 if (pte_write(pteval)) {
 598                         writable = true;
 599                 } else {
 600                         if (PageSwapCache(page) &&
 601                             !reuse_swap_page(page, NULL)) {
 602                                 unlock_page(page);
 603                                 result = SCAN_SWAP_CACHE_PAGE;
 604                                 goto out;
 605                         }
 606                         /*
 607                          * Page is not in the swap cache. It can be collapsed
 608                          * into a THP.
 609                          */
 610                 }
 611
 612                 /*
 613                  * Isolate the page to avoid collapsing an hugepage
 614                  * currently in use by the VM.
 615                  */
 616                 if (isolate_lru_page(page)) {
 617                         unlock_page(page);
 618                         result = SCAN_DEL_PAGE_LRU;
 619                         goto out;
 620                 }
 621                 inc_node_page_state(page,
 622                                 NR_ISOLATED_ANON + page_is_file_cache(page));
 623                 VM_BUG_ON_PAGE(!PageLocked(page), page);
 624                 VM_BUG_ON_PAGE(PageLRU(page), page);
 625
 626                 /* There should be enough young pte to collapse the page */
 627                 if (pte_young(pteval) ||
 628                     page_is_young(page) || PageReferenced(page) ||
 629                     mmu_notifier_test_young(vma->vm_mm, address))
 630                         referenced++;
 631         }
 632
 633         if (unlikely(!writable)) {
 634                 result = SCAN_PAGE_RO;
 635         } else if (unlikely(!referenced)) {
 636                 result = SCAN_LACK_REFERENCED_PAGE;
 637         } else {
 638                 result = SCAN_SUCCEED;
 639                 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 640                                                     referenced, writable, result);
 641                 return 1;
 642         }
 643 out:
 644         release_pte_pages(pte, _pte);
 645         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 646                                             referenced, writable, result);
 647         return 0;
 648 }
 649
 650 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
 651                                       struct vm_area_struct *vma,
 652                                       unsigned long address,
 653                                       spinlock_t *ptl)
 654 {
 655         pte_t *_pte;
 656         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 657                                 _pte++, page++, address += PAGE_SIZE) {
 658                 pte_t pteval = *_pte;
 659                 struct page *src_page;
 660
 661                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
 662                         clear_user_highpage(page, address);
 663                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
 664                         if (is_zero_pfn(pte_pfn(pteval))) {
 665                                 /*
 666                                  * ptl mostly unnecessary.
 667                                  */
 668                                 spin_lock(ptl);
 669                                 /*
 670                                  * paravirt calls inside pte_clear here are
 671                                  * superfluous.
 672                                  */
 673                                 pte_clear(vma->vm_mm, address, _pte);
 674                                 spin_unlock(ptl);
 675                         }
 676                 } else {
 677                         src_page = pte_page(pteval);
 678                         copy_user_highpage(page, src_page, address, vma);
 679                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
 680                         release_pte_page(src_page);
 681                         /*
 682                          * ptl mostly unnecessary, but preempt has to
 683                          * be disabled to update the per-cpu stats
 684                          * inside page_remove_rmap().
 685                          */
 686                         spin_lock(ptl);
 687                         /*
 688                          * paravirt calls inside pte_clear here are
 689                          * superfluous.
 690                          */
 691                         pte_clear(vma->vm_mm, address, _pte);
 692                         page_remove_rmap(src_page, false);
 693                         spin_unlock(ptl);
 694                         free_page_and_swap_cache(src_page);
 695                 }
 696         }
 697 }
 698
 699 static void khugepaged_alloc_sleep(void)
 700 {
 701         DEFINE_WAIT(wait);
 702
 703         add_wait_queue(&khugepaged_wait, &wait);
 704         freezable_schedule_timeout_interruptible(
 705                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
 706         remove_wait_queue(&khugepaged_wait, &wait);
 707 }
 708
 709 static int khugepaged_node_load[MAX_NUMNODES];
 710
 711 static bool khugepaged_scan_abort(int nid)
 712 {
 713         int i;
 714
 715         /*
 716          * If node_reclaim_mode is disabled, then no extra effort is made to
 717          * allocate memory locally.
 718          */
 719         if (!node_reclaim_mode)
 720                 return false;
 721
 722         /* If there is a count for this node already, it must be acceptable */
 723         if (khugepaged_node_load[nid])
 724                 return false;
 725
 726         for (i = 0; i < MAX_NUMNODES; i++) {
 727                 if (!khugepaged_node_load[i])
 728                         continue;
 729                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
 730                         return true;
 731         }
 732         return false;
 733 }
 734
 735 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
 736 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
 737 {
 738         return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
 739 }
 740
 741 #ifdef CONFIG_NUMA
 742 static int khugepaged_find_target_node(void)
 743 {
 744         static int last_khugepaged_target_node = NUMA_NO_NODE;
 745         int nid, target_node = 0, max_value = 0;
 746
 747         /* find first node with max normal pages hit */
 748         for (nid = 0; nid < MAX_NUMNODES; nid++)
 749                 if (khugepaged_node_load[nid] > max_value) {
 750                         max_value = khugepaged_node_load[nid];
 751                         target_node = nid;
 752                 }
 753
 754         /* do some balance if several nodes have the same hit record */
 755         if (target_node <= last_khugepaged_target_node)
 756                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
 757                                 nid++)
 758                         if (max_value == khugepaged_node_load[nid]) {
 759                                 target_node = nid;
 760                                 break;
 761                         }
 762
 763         last_khugepaged_target_node = target_node;
 764         return target_node;
 765 }
 766
 767 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 768 {
 769         if (IS_ERR(*hpage)) {
 770                 if (!*wait)
 771                         return false;
 772
 773                 *wait = false;
 774                 *hpage = NULL;
 775                 khugepaged_alloc_sleep();
 776         } else if (*hpage) {
 777                 put_page(*hpage);
 778                 *hpage = NULL;
 779         }
 780
 781         return true;
 782 }
 783
 784 static struct page *
 785 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 786 {
 787         VM_BUG_ON_PAGE(*hpage, *hpage);
 788
 789         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
 790         if (unlikely(!*hpage)) {
 791                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 792                 *hpage = ERR_PTR(-ENOMEM);
 793                 return NULL;
 794         }
 795
 796         prep_transhuge_page(*hpage);
 797         count_vm_event(THP_COLLAPSE_ALLOC);
 798         return *hpage;
 799 }
 800 #else
 801 static int khugepaged_find_target_node(void)
 802 {
 803         return 0;
 804 }
 805
 806 static inline struct page *alloc_khugepaged_hugepage(void)
 807 {
 808         struct page *page;
 809
 810         page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
 811                            HPAGE_PMD_ORDER);
 812         if (page)
 813                 prep_transhuge_page(page);
 814         return page;
 815 }
 816
 817 static struct page *khugepaged_alloc_hugepage(bool *wait)
 818 {
 819         struct page *hpage;
 820
 821         do {
 822                 hpage = alloc_khugepaged_hugepage();
 823                 if (!hpage) {
 824                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 825                         if (!*wait)
 826                                 return NULL;
 827
 828                         *wait = false;
 829                         khugepaged_alloc_sleep();
 830                 } else
 831                         count_vm_event(THP_COLLAPSE_ALLOC);
 832         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
 833
 834         return hpage;
 835 }
 836
 837 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 838 {
 839         /*
 840          * If the hpage allocated earlier was briefly exposed in page cache
 841          * before collapse_file() failed, it is possible that racing lookups
 842          * have not yet completed, and would then be unpleasantly surprised by
 843          * finding the hpage reused for the same mapping at a different offset.
 844          * Just release the previous allocation if there is any danger of that.
 845          */
 846         if (*hpage && page_count(*hpage) > 1) {
 847                 put_page(*hpage);
 848                 *hpage = NULL;
 849         }
 850
 851         if (!*hpage)
 852                 *hpage = khugepaged_alloc_hugepage(wait);
 853
 854         if (unlikely(!*hpage))
 855                 return false;
 856
 857         return true;
 858 }
 859
 860 static struct page *
 861 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 862 {
 863         VM_BUG_ON(!*hpage);
 864
 865         return  *hpage;
 866 }
 867 #endif
 868
 869 /*
 870  * If mmap_sem temporarily dropped, revalidate vma
 871  * before taking mmap_sem.
 872  * Return 0 if succeeds, otherwise return none-zero
 873  * value (scan code).
 874  */
 875
 876 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
 877                 struct vm_area_struct **vmap)
 878 {
 879         struct vm_area_struct *vma;
 880         unsigned long hstart, hend;
 881
 882         if (unlikely(khugepaged_test_exit(mm)))
 883                 return SCAN_ANY_PROCESS;
 884
 885         *vmap = vma = find_vma(mm, address);
 886         if (!vma)
 887                 return SCAN_VMA_NULL;
 888
 889         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 890         hend = vma->vm_end & HPAGE_PMD_MASK;
 891         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
 892                 return SCAN_ADDRESS_RANGE;
 893         if (!hugepage_vma_check(vma, vma->vm_flags))
 894                 return SCAN_VMA_CHECK;
 895         return 0;
 896 }
 897
 898 /*
 899  * Bring missing pages in from swap, to complete THP collapse.
 900  * Only done if khugepaged_scan_pmd believes it is worthwhile.
 901  *
 902  * Called and returns without pte mapped or spinlocks held,
 903  * but with mmap_sem held to protect against vma changes.
 904  */
 905
 906 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
 907                                         struct vm_area_struct *vma,
 908                                         unsigned long address, pmd_t *pmd,
 909                                         int referenced)
 910 {
 911         int swapped_in = 0;
 912         vm_fault_t ret = 0;
 913         struct vm_fault vmf = {
 914                 .vma = vma,
 915                 .address = address,
 916                 .flags = FAULT_FLAG_ALLOW_RETRY,
 917                 .pmd = pmd,
 918                 .pgoff = linear_page_index(vma, address),
 919         };
 920
 921         /* we only decide to swapin, if there is enough young ptes */
 922         if (referenced < HPAGE_PMD_NR/2) {
 923                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 924                 return false;
 925         }
 926         vmf.pte = pte_offset_map(pmd, address);
 927         for (; vmf.address < address + HPAGE_PMD_NR*PAGE_SIZE;
 928                         vmf.pte++, vmf.address += PAGE_SIZE) {
 929                 vmf.orig_pte = *vmf.pte;
 930                 if (!is_swap_pte(vmf.orig_pte))
 931                         continue;
 932                 swapped_in++;
 933                 ret = do_swap_page(&vmf);
 934
 935                 /* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */
 936                 if (ret & VM_FAULT_RETRY) {
 937                         down_read(&mm->mmap_sem);
 938                         if (hugepage_vma_revalidate(mm, address, &vmf.vma)) {
 939                                 /* vma is no longer available, don't continue to swapin */
 940                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 941                                 return false;
 942                         }
 943                         /* check if the pmd is still valid */
 944                         if (mm_find_pmd(mm, address) != pmd) {
 945                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 946                                 return false;
 947                         }
 948                 }
 949                 if (ret & VM_FAULT_ERROR) {
 950                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 951                         return false;
 952                 }
 953                 /* pte is unmapped now, we need to map it */
 954                 vmf.pte = pte_offset_map(pmd, vmf.address);
 955         }
 956         vmf.pte--;
 957         pte_unmap(vmf.pte);
 958         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
 959         return true;
 960 }
 961
 962 static void collapse_huge_page(struct mm_struct *mm,
 963                                    unsigned long address,
 964                                    struct page **hpage,
 965                                    int node, int referenced)
 966 {
 967         pmd_t *pmd, _pmd;
 968         pte_t *pte;
 969         pgtable_t pgtable;
 970         struct page *new_page;
 971         spinlock_t *pmd_ptl, *pte_ptl;
 972         int isolated = 0, result = 0;
 973         struct mem_cgroup *memcg;
 974         struct vm_area_struct *vma;
 975         unsigned long mmun_start;       /* For mmu_notifiers */
 976         unsigned long mmun_end;         /* For mmu_notifiers */
 977         gfp_t gfp;
 978
 979         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 980
 981         /* Only allocate from the target node */
 982         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
 983
 984         /*
 985          * Before allocating the hugepage, release the mmap_sem read lock.
 986          * The allocation can take potentially a long time if it involves
 987          * sync compaction, and we do not need to hold the mmap_sem during
 988          * that. We will recheck the vma after taking it again in write mode.
 989          */
 990         up_read(&mm->mmap_sem);
 991         new_page = khugepaged_alloc_page(hpage, gfp, node);
 992         if (!new_page) {
 993                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
 994                 goto out_nolock;
 995         }
 996
 997         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
 998                 result = SCAN_CGROUP_CHARGE_FAIL;
 999                 goto out_nolock;
1000         }
1001
1002         down_read(&mm->mmap_sem);
1003         result = hugepage_vma_revalidate(mm, address, &vma);
1004         if (result) {
1005                 mem_cgroup_cancel_charge(new_page, memcg, true);
1006                 up_read(&mm->mmap_sem);
1007                 goto out_nolock;
1008         }
1009
1010         pmd = mm_find_pmd(mm, address);
1011         if (!pmd) {
1012                 result = SCAN_PMD_NULL;
1013                 mem_cgroup_cancel_charge(new_page, memcg, true);
1014                 up_read(&mm->mmap_sem);
1015                 goto out_nolock;
1016         }
1017
1018         /*
1019          * __collapse_huge_page_swapin always returns with mmap_sem locked.
1020          * If it fails, we release mmap_sem and jump out_nolock.
1021          * Continuing to collapse causes inconsistency.
1022          */
1023         if (!__collapse_huge_page_swapin(mm, vma, address, pmd, referenced)) {
1024                 mem_cgroup_cancel_charge(new_page, memcg, true);
1025                 up_read(&mm->mmap_sem);
1026                 goto out_nolock;
1027         }
1028
1029         up_read(&mm->mmap_sem);
1030         /*
1031          * Prevent all access to pagetables with the exception of
1032          * gup_fast later handled by the ptep_clear_flush and the VM
1033          * handled by the anon_vma lock + PG_lock.
1034          */
1035         down_write(&mm->mmap_sem);
1036         result = hugepage_vma_revalidate(mm, address, &vma);
1037         if (result)
1038                 goto out;
1039         /* check if the pmd is still valid */
1040         if (mm_find_pmd(mm, address) != pmd)
1041                 goto out;
1042
1043         anon_vma_lock_write(vma->anon_vma);
1044
1045         pte = pte_offset_map(pmd, address);
1046         pte_ptl = pte_lockptr(mm, pmd);
1047
1048         mmun_start = address;
1049         mmun_end   = address + HPAGE_PMD_SIZE;
1050         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1051         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1052         /*
1053          * After this gup_fast can't run anymore. This also removes
1054          * any huge TLB entry from the CPU so we won't allow
1055          * huge and small TLB entries for the same virtual address
1056          * to avoid the risk of CPU bugs in that area.
1057          */
1058         _pmd = pmdp_collapse_flush(vma, address, pmd);
1059         spin_unlock(pmd_ptl);
1060         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1061         tlb_remove_table_sync_one();
1062
1063         spin_lock(pte_ptl);
1064         isolated = __collapse_huge_page_isolate(vma, address, pte);
1065         spin_unlock(pte_ptl);
1066
1067         if (unlikely(!isolated)) {
1068                 pte_unmap(pte);
1069                 spin_lock(pmd_ptl);
1070                 BUG_ON(!pmd_none(*pmd));
1071                 /*
1072                  * We can only use set_pmd_at when establishing
1073                  * hugepmds and never for establishing regular pmds that
1074                  * points to regular pagetables. Use pmd_populate for that
1075                  */
1076                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1077                 spin_unlock(pmd_ptl);
1078                 anon_vma_unlock_write(vma->anon_vma);
1079                 result = SCAN_FAIL;
1080                 goto out;
1081         }
1082
1083         /*
1084          * All pages are isolated and locked so anon_vma rmap
1085          * can't run anymore.
1086          */
1087         anon_vma_unlock_write(vma->anon_vma);
1088
1089         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
1090         pte_unmap(pte);
1091         __SetPageUptodate(new_page);
1092         pgtable = pmd_pgtable(_pmd);
1093
1094         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1095         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1096
1097         /*
1098          * spin_lock() below is not the equivalent of smp_wmb(), so
1099          * this is needed to avoid the copy_huge_page writes to become
1100          * visible after the set_pmd_at() write.
1101          */
1102         smp_wmb();
1103
1104         spin_lock(pmd_ptl);
1105         BUG_ON(!pmd_none(*pmd));
1106         page_add_new_anon_rmap(new_page, vma, address, true);
1107         mem_cgroup_commit_charge(new_page, memcg, false, true);
1108         lru_cache_add_active_or_unevictable(new_page, vma);
1109         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1110         set_pmd_at(mm, address, pmd, _pmd);
1111         update_mmu_cache_pmd(vma, address, pmd);
1112         spin_unlock(pmd_ptl);
1113
1114         *hpage = NULL;
1115
1116         khugepaged_pages_collapsed++;
1117         result = SCAN_SUCCEED;
1118 out_up_write:
1119         up_write(&mm->mmap_sem);
1120 out_nolock:
1121         trace_mm_collapse_huge_page(mm, isolated, result);
1122         return;
1123 out:
1124         mem_cgroup_cancel_charge(new_page, memcg, true);
1125         goto out_up_write;
1126 }
1127
1128 static int khugepaged_scan_pmd(struct mm_struct *mm,
1129                                struct vm_area_struct *vma,
1130                                unsigned long address,
1131                                struct page **hpage)
1132 {
1133         pmd_t *pmd;
1134         pte_t *pte, *_pte;
1135         int ret = 0, none_or_zero = 0, result = 0, referenced = 0;
1136         struct page *page = NULL;
1137         unsigned long _address;
1138         spinlock_t *ptl;
1139         int node = NUMA_NO_NODE, unmapped = 0;
1140         bool writable = false;
1141
1142         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1143
1144         pmd = mm_find_pmd(mm, address);
1145         if (!pmd) {
1146                 result = SCAN_PMD_NULL;
1147                 goto out;
1148         }
1149
1150         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1151         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1152         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1153              _pte++, _address += PAGE_SIZE) {
1154                 pte_t pteval = *_pte;
1155                 if (is_swap_pte(pteval)) {
1156                         if (++unmapped <= khugepaged_max_ptes_swap) {
1157                                 continue;
1158                         } else {
1159                                 result = SCAN_EXCEED_SWAP_PTE;
1160                                 goto out_unmap;
1161                         }
1162                 }
1163                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1164                         if (!userfaultfd_armed(vma) &&
1165                             ++none_or_zero <= khugepaged_max_ptes_none) {
1166                                 continue;
1167                         } else {
1168                                 result = SCAN_EXCEED_NONE_PTE;
1169                                 goto out_unmap;
1170                         }
1171                 }
1172                 if (!pte_present(pteval)) {
1173                         result = SCAN_PTE_NON_PRESENT;
1174                         goto out_unmap;
1175                 }
1176                 if (pte_write(pteval))
1177                         writable = true;
1178
1179                 page = vm_normal_page(vma, _address, pteval);
1180                 if (unlikely(!page)) {
1181                         result = SCAN_PAGE_NULL;
1182                         goto out_unmap;
1183                 }
1184
1185                 /* TODO: teach khugepaged to collapse THP mapped with pte */
1186                 if (PageCompound(page)) {
1187                         result = SCAN_PAGE_COMPOUND;
1188                         goto out_unmap;
1189                 }
1190
1191                 /*
1192                  * Record which node the original page is from and save this
1193                  * information to khugepaged_node_load[].
1194                  * Khupaged will allocate hugepage from the node has the max
1195                  * hit record.
1196                  */
1197                 node = page_to_nid(page);
1198                 if (khugepaged_scan_abort(node)) {
1199                         result = SCAN_SCAN_ABORT;
1200                         goto out_unmap;
1201                 }
1202                 khugepaged_node_load[node]++;
1203                 if (!PageLRU(page)) {
1204                         result = SCAN_PAGE_LRU;
1205                         goto out_unmap;
1206                 }
1207                 if (PageLocked(page)) {
1208                         result = SCAN_PAGE_LOCK;
1209                         goto out_unmap;
1210                 }
1211                 if (!PageAnon(page)) {
1212                         result = SCAN_PAGE_ANON;
1213                         goto out_unmap;
1214                 }
1215
1216                 /*
1217                  * cannot use mapcount: can't collapse if there's a gup pin.
1218                  * The page must only be referenced by the scanned process
1219                  * and page swap cache.
1220                  */
1221                 if (page_count(page) != 1 + PageSwapCache(page)) {
1222                         result = SCAN_PAGE_COUNT;
1223                         goto out_unmap;
1224                 }
1225                 if (pte_young(pteval) ||
1226                     page_is_young(page) || PageReferenced(page) ||
1227                     mmu_notifier_test_young(vma->vm_mm, address))
1228                         referenced++;
1229         }
1230         if (writable) {
1231                 if (referenced) {
1232                         result = SCAN_SUCCEED;
1233                         ret = 1;
1234                 } else {
1235                         result = SCAN_LACK_REFERENCED_PAGE;
1236                 }
1237         } else {
1238                 result = SCAN_PAGE_RO;
1239         }
1240 out_unmap:
1241         pte_unmap_unlock(pte, ptl);
1242         if (ret) {
1243                 node = khugepaged_find_target_node();
1244                 /* collapse_huge_page will return with the mmap_sem released */
1245                 collapse_huge_page(mm, address, hpage, node, referenced);
1246         }
1247 out:
1248         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1249                                      none_or_zero, result, unmapped);
1250         return ret;
1251 }
1252
1253 static void collect_mm_slot(struct mm_slot *mm_slot)
1254 {
1255         struct mm_struct *mm = mm_slot->mm;
1256
1257         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
1258
1259         if (khugepaged_test_exit(mm)) {
1260                 /* free mm_slot */
1261                 hash_del(&mm_slot->hash);
1262                 list_del(&mm_slot->mm_node);
1263
1264                 /*
1265                  * Not strictly needed because the mm exited already.
1266                  *
1267                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1268                  */
1269
1270                 /* khugepaged_mm_lock actually not necessary for the below */
1271                 free_mm_slot(mm_slot);
1272                 mmdrop(mm);
1273         }
1274 }
1275
1276 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
1277 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1278 {
1279         struct vm_area_struct *vma;
1280         struct mm_struct *mm;
1281         unsigned long addr;
1282         pmd_t *pmd, _pmd;
1283
1284         i_mmap_lock_write(mapping);
1285         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1286                 /* probably overkill */
1287                 if (vma->anon_vma)
1288                         continue;
1289                 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1290                 if (addr & ~HPAGE_PMD_MASK)
1291                         continue;
1292                 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1293                         continue;
1294                 mm = vma->vm_mm;
1295                 pmd = mm_find_pmd(mm, addr);
1296                 if (!pmd)
1297                         continue;
1298                 /*
1299                  * We need exclusive mmap_sem to retract page table.
1300                  * If trylock fails we would end up with pte-mapped THP after
1301                  * re-fault. Not ideal, but it's more important to not disturb
1302                  * the system too much.
1303                  */
1304                 if (down_write_trylock(&mm->mmap_sem)) {
1305                         if (!khugepaged_test_exit(mm)) {
1306                                 spinlock_t *ptl;
1307                                 unsigned long end = addr + HPAGE_PMD_SIZE;
1308
1309                                 mmu_notifier_invalidate_range_start(mm, addr,
1310                                                                     end);
1311                                 ptl = pmd_lock(mm, pmd);
1312                                 /* assume page table is clear */
1313                                 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1314                                 spin_unlock(ptl);
1315                                 mm_dec_nr_ptes(mm);
1316                                 tlb_remove_table_sync_one();
1317                                 pte_free(mm, pmd_pgtable(_pmd));
1318                                 mmu_notifier_invalidate_range_end(mm, addr,
1319                                                                   end);
1320                         }
1321                         up_write(&mm->mmap_sem);
1322                 }
1323         }
1324         i_mmap_unlock_write(mapping);
1325 }
1326
1327 /**
1328  * collapse_shmem - collapse small tmpfs/shmem pages into huge one.
1329  *
1330  * Basic scheme is simple, details are more complex:
1331  *  - allocate and lock a new huge page;
1332  *  - scan over radix tree replacing old pages the new one
1333  *    + swap in pages if necessary;
1334  *    + fill in gaps;
1335  *    + keep old pages around in case if rollback is required;
1336  *  - if replacing succeed:
1337  *    + copy data over;
1338  *    + free old pages;
1339  *    + unlock huge page;
1340  *  - if replacing failed;
1341  *    + put all pages back and unfreeze them;
1342  *    + restore gaps in the radix-tree;
1343  *    + unlock and free huge page;
1344  */
1345 static void collapse_shmem(struct mm_struct *mm,
1346                 struct address_space *mapping, pgoff_t start,
1347                 struct page **hpage, int node)
1348 {
1349         gfp_t gfp;
1350         struct page *page, *new_page, *tmp;
1351         struct mem_cgroup *memcg;
1352         pgoff_t index, end = start + HPAGE_PMD_NR;
1353         LIST_HEAD(pagelist);
1354         struct radix_tree_iter iter;
1355         void **slot;
1356         int nr_none = 0, result = SCAN_SUCCEED;
1357
1358         VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1359
1360         /* Only allocate from the target node */
1361         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1362
1363         new_page = khugepaged_alloc_page(hpage, gfp, node);
1364         if (!new_page) {
1365                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1366                 goto out;
1367         }
1368
1369         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
1370                 result = SCAN_CGROUP_CHARGE_FAIL;
1371                 goto out;
1372         }
1373
1374         __SetPageLocked(new_page);
1375         __SetPageSwapBacked(new_page);
1376         new_page->index = start;
1377         new_page->mapping = mapping;
1378
1379         /*
1380          * At this point the new_page is locked and not up-to-date.
1381          * It's safe to insert it into the page cache, because nobody would
1382          * be able to map it or use it in another way until we unlock it.
1383          */
1384
1385         index = start;
1386         xa_lock_irq(&mapping->i_pages);
1387         radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start) {
1388                 int n = min(iter.index, end) - index;
1389
1390                 /*
1391                  * Stop if extent has been hole-punched, and is now completely
1392                  * empty (the more obvious i_size_read() check would take an
1393                  * irq-unsafe seqlock on 32-bit).
1394                  */
1395                 if (n >= HPAGE_PMD_NR) {
1396                         result = SCAN_TRUNCATED;
1397                         goto tree_locked;
1398                 }
1399
1400                 /*
1401                  * Handle holes in the radix tree: charge it from shmem and
1402                  * insert relevant subpage of new_page into the radix-tree.
1403                  */
1404                 if (n && !shmem_charge(mapping->host, n)) {
1405                         result = SCAN_FAIL;
1406                         goto tree_locked;
1407                 }
1408                 for (; index < min(iter.index, end); index++) {
1409                         radix_tree_insert(&mapping->i_pages, index,
1410                                         new_page + (index % HPAGE_PMD_NR));
1411                 }
1412                 nr_none += n;
1413
1414                 /* We are done. */
1415                 if (index >= end)
1416                         break;
1417
1418                 page = radix_tree_deref_slot_protected(slot,
1419                                 &mapping->i_pages.xa_lock);
1420                 if (radix_tree_exceptional_entry(page) || !PageUptodate(page)) {
1421                         xa_unlock_irq(&mapping->i_pages);
1422                         /* swap in or instantiate fallocated page */
1423                         if (shmem_getpage(mapping->host, index, &page,
1424                                                 SGP_NOHUGE)) {
1425                                 result = SCAN_FAIL;
1426                                 goto tree_unlocked;
1427                         }
1428                 } else if (trylock_page(page)) {
1429                         get_page(page);
1430                         xa_unlock_irq(&mapping->i_pages);
1431                 } else {
1432                         result = SCAN_PAGE_LOCK;
1433                         goto tree_locked;
1434                 }
1435
1436                 /*
1437                  * The page must be locked, so we can drop the i_pages lock
1438                  * without racing with truncate.
1439                  */
1440                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1441                 VM_BUG_ON_PAGE(!PageUptodate(page), page);
1442
1443                 /*
1444                  * If file was truncated then extended, or hole-punched, before
1445                  * we locked the first page, then a THP might be there already.
1446                  */
1447                 if (PageTransCompound(page)) {
1448                         result = SCAN_PAGE_COMPOUND;
1449                         goto out_unlock;
1450                 }
1451
1452                 if (page_mapping(page) != mapping) {
1453                         result = SCAN_TRUNCATED;
1454                         goto out_unlock;
1455                 }
1456
1457                 if (isolate_lru_page(page)) {
1458                         result = SCAN_DEL_PAGE_LRU;
1459                         goto out_unlock;
1460                 }
1461
1462                 if (page_mapped(page))
1463                         unmap_mapping_pages(mapping, index, 1, false);
1464
1465                 xa_lock_irq(&mapping->i_pages);
1466
1467                 slot = radix_tree_lookup_slot(&mapping->i_pages, index);
1468                 VM_BUG_ON_PAGE(page != radix_tree_deref_slot_protected(slot,
1469                                         &mapping->i_pages.xa_lock), page);
1470                 VM_BUG_ON_PAGE(page_mapped(page), page);
1471
1472                 /*
1473                  * The page is expected to have page_count() == 3:
1474                  *  - we hold a pin on it;
1475                  *  - one reference from radix tree;
1476                  *  - one from isolate_lru_page;
1477                  */
1478                 if (!page_ref_freeze(page, 3)) {
1479                         result = SCAN_PAGE_COUNT;
1480                         xa_unlock_irq(&mapping->i_pages);
1481                         putback_lru_page(page);
1482                         goto out_unlock;
1483                 }
1484
1485                 /*
1486                  * Add the page to the list to be able to undo the collapse if
1487                  * something go wrong.
1488                  */
1489                 list_add_tail(&page->lru, &pagelist);
1490
1491                 /* Finally, replace with the new page. */
1492                 radix_tree_replace_slot(&mapping->i_pages, slot,
1493                                 new_page + (index % HPAGE_PMD_NR));
1494
1495                 slot = radix_tree_iter_resume(slot, &iter);
1496                 index++;
1497                 continue;
1498 out_unlock:
1499                 unlock_page(page);
1500                 put_page(page);
1501                 goto tree_unlocked;
1502         }
1503
1504         /*
1505          * Handle hole in radix tree at the end of the range.
1506          * This code only triggers if there's nothing in radix tree
1507          * beyond 'end'.
1508          */
1509         if (index < end) {
1510                 int n = end - index;
1511
1512                 /* Stop if extent has been truncated, and is now empty */
1513                 if (n >= HPAGE_PMD_NR) {
1514                         result = SCAN_TRUNCATED;
1515                         goto tree_locked;
1516                 }
1517                 if (!shmem_charge(mapping->host, n)) {
1518                         result = SCAN_FAIL;
1519                         goto tree_locked;
1520                 }
1521                 for (; index < end; index++) {
1522                         radix_tree_insert(&mapping->i_pages, index,
1523                                         new_page + (index % HPAGE_PMD_NR));
1524                 }
1525                 nr_none += n;
1526         }
1527
1528         __inc_node_page_state(new_page, NR_SHMEM_THPS);
1529         if (nr_none) {
1530                 struct zone *zone = page_zone(new_page);
1531
1532                 __mod_node_page_state(zone->zone_pgdat, NR_FILE_PAGES, nr_none);
1533                 __mod_node_page_state(zone->zone_pgdat, NR_SHMEM, nr_none);
1534         }
1535
1536 tree_locked:
1537         xa_unlock_irq(&mapping->i_pages);
1538 tree_unlocked:
1539
1540         if (result == SCAN_SUCCEED) {
1541                 /*
1542                  * Replacing old pages with new one has succeed, now we need to
1543                  * copy the content and free old pages.
1544                  */
1545                 index = start;
1546                 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1547                         while (index < page->index) {
1548                                 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1549                                 index++;
1550                         }
1551                         copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1552                                         page);
1553                         list_del(&page->lru);
1554                         page->mapping = NULL;
1555                         page_ref_unfreeze(page, 1);
1556                         ClearPageActive(page);
1557                         ClearPageUnevictable(page);
1558                         unlock_page(page);
1559                         put_page(page);
1560                         index++;
1561                 }
1562                 while (index < end) {
1563                         clear_highpage(new_page + (index % HPAGE_PMD_NR));
1564                         index++;
1565                 }
1566
1567                 SetPageUptodate(new_page);
1568                 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1569                 set_page_dirty(new_page);
1570                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1571                 lru_cache_add_anon(new_page);
1572
1573                 /*
1574                  * Remove pte page tables, so we can re-fault the page as huge.
1575                  */
1576                 retract_page_tables(mapping, start);
1577                 *hpage = NULL;
1578
1579                 khugepaged_pages_collapsed++;
1580         } else {
1581                 /* Something went wrong: rollback changes to the radix-tree */
1582                 xa_lock_irq(&mapping->i_pages);
1583                 mapping->nrpages -= nr_none;
1584                 shmem_uncharge(mapping->host, nr_none);
1585
1586                 radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start) {
1587                         if (iter.index >= end)
1588                                 break;
1589                         page = list_first_entry_or_null(&pagelist,
1590                                         struct page, lru);
1591                         if (!page || iter.index < page->index) {
1592                                 if (!nr_none)
1593                                         break;
1594                                 nr_none--;
1595                                 /* Put holes back where they were */
1596                                 radix_tree_delete(&mapping->i_pages, iter.index);
1597                                 continue;
1598                         }
1599
1600                         VM_BUG_ON_PAGE(page->index != iter.index, page);
1601
1602                         /* Unfreeze the page. */
1603                         list_del(&page->lru);
1604                         page_ref_unfreeze(page, 2);
1605                         radix_tree_replace_slot(&mapping->i_pages, slot, page);
1606                         slot = radix_tree_iter_resume(slot, &iter);
1607                         xa_unlock_irq(&mapping->i_pages);
1608                         unlock_page(page);
1609                         putback_lru_page(page);
1610                         xa_lock_irq(&mapping->i_pages);
1611                 }
1612                 VM_BUG_ON(nr_none);
1613                 xa_unlock_irq(&mapping->i_pages);
1614
1615                 mem_cgroup_cancel_charge(new_page, memcg, true);
1616                 new_page->mapping = NULL;
1617         }
1618
1619         unlock_page(new_page);
1620 out:
1621         VM_BUG_ON(!list_empty(&pagelist));
1622         /* TODO: tracepoints */
1623 }
1624
1625 static void khugepaged_scan_shmem(struct mm_struct *mm,
1626                 struct address_space *mapping,
1627                 pgoff_t start, struct page **hpage)
1628 {
1629         struct page *page = NULL;
1630         struct radix_tree_iter iter;
1631         void **slot;
1632         int present, swap;
1633         int node = NUMA_NO_NODE;
1634         int result = SCAN_SUCCEED;
1635
1636         present = 0;
1637         swap = 0;
1638         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1639         rcu_read_lock();
1640         radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start) {
1641                 if (iter.index >= start + HPAGE_PMD_NR)
1642                         break;
1643
1644                 page = radix_tree_deref_slot(slot);
1645                 if (radix_tree_deref_retry(page)) {
1646                         slot = radix_tree_iter_retry(&iter);
1647                         continue;
1648                 }
1649
1650                 if (radix_tree_exception(page)) {
1651                         if (++swap > khugepaged_max_ptes_swap) {
1652                                 result = SCAN_EXCEED_SWAP_PTE;
1653                                 break;
1654                         }
1655                         continue;
1656                 }
1657
1658                 if (PageTransCompound(page)) {
1659                         result = SCAN_PAGE_COMPOUND;
1660                         break;
1661                 }
1662
1663                 node = page_to_nid(page);
1664                 if (khugepaged_scan_abort(node)) {
1665                         result = SCAN_SCAN_ABORT;
1666                         break;
1667                 }
1668                 khugepaged_node_load[node]++;
1669
1670                 if (!PageLRU(page)) {
1671                         result = SCAN_PAGE_LRU;
1672                         break;
1673                 }
1674
1675                 if (page_count(page) != 1 + page_mapcount(page)) {
1676                         result = SCAN_PAGE_COUNT;
1677                         break;
1678                 }
1679
1680                 /*
1681                  * We probably should check if the page is referenced here, but
1682                  * nobody would transfer pte_young() to PageReferenced() for us.
1683                  * And rmap walk here is just too costly...
1684                  */
1685
1686                 present++;
1687
1688                 if (need_resched()) {
1689                         slot = radix_tree_iter_resume(slot, &iter);
1690                         cond_resched_rcu();
1691                 }
1692         }
1693         rcu_read_unlock();
1694
1695         if (result == SCAN_SUCCEED) {
1696                 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
1697                         result = SCAN_EXCEED_NONE_PTE;
1698                 } else {
1699                         node = khugepaged_find_target_node();
1700                         collapse_shmem(mm, mapping, start, hpage, node);
1701                 }
1702         }
1703
1704         /* TODO: tracepoints */
1705 }
1706 #else
1707 static void khugepaged_scan_shmem(struct mm_struct *mm,
1708                 struct address_space *mapping,
1709                 pgoff_t start, struct page **hpage)
1710 {
1711         BUILD_BUG();
1712 }
1713 #endif
1714
1715 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
1716                                             struct page **hpage)
1717         __releases(&khugepaged_mm_lock)
1718         __acquires(&khugepaged_mm_lock)
1719 {
1720         struct mm_slot *mm_slot;
1721         struct mm_struct *mm;
1722         struct vm_area_struct *vma;
1723         int progress = 0;
1724
1725         VM_BUG_ON(!pages);
1726         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
1727
1728         if (khugepaged_scan.mm_slot)
1729                 mm_slot = khugepaged_scan.mm_slot;
1730         else {
1731                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
1732                                      struct mm_slot, mm_node);
1733                 khugepaged_scan.address = 0;
1734                 khugepaged_scan.mm_slot = mm_slot;
1735         }
1736         spin_unlock(&khugepaged_mm_lock);
1737
1738         mm = mm_slot->mm;
1739         /*
1740          * Don't wait for semaphore (to avoid long wait times).  Just move to
1741          * the next mm on the list.
1742          */
1743         vma = NULL;
1744         if (unlikely(!down_read_trylock(&mm->mmap_sem)))
1745                 goto breakouterloop_mmap_sem;
1746         if (likely(!khugepaged_test_exit(mm)))
1747                 vma = find_vma(mm, khugepaged_scan.address);
1748
1749         progress++;
1750         for (; vma; vma = vma->vm_next) {
1751                 unsigned long hstart, hend;
1752
1753                 cond_resched();
1754                 if (unlikely(khugepaged_test_exit(mm))) {
1755                         progress++;
1756                         break;
1757                 }
1758                 if (!hugepage_vma_check(vma, vma->vm_flags)) {
1759 skip:
1760                         progress++;
1761                         continue;
1762                 }
1763                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1764                 hend = vma->vm_end & HPAGE_PMD_MASK;
1765                 if (hstart >= hend)
1766                         goto skip;
1767                 if (khugepaged_scan.address > hend)
1768                         goto skip;
1769                 if (khugepaged_scan.address < hstart)
1770                         khugepaged_scan.address = hstart;
1771                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
1772
1773                 while (khugepaged_scan.address < hend) {
1774                         int ret;
1775                         cond_resched();
1776                         if (unlikely(khugepaged_test_exit(mm)))
1777                                 goto breakouterloop;
1778
1779                         VM_BUG_ON(khugepaged_scan.address < hstart ||
1780                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
1781                                   hend);
1782                         if (shmem_file(vma->vm_file)) {
1783                                 struct file *file;
1784                                 pgoff_t pgoff = linear_page_index(vma,
1785                                                 khugepaged_scan.address);
1786                                 if (!shmem_huge_enabled(vma))
1787                                         goto skip;
1788                                 file = get_file(vma->vm_file);
1789                                 up_read(&mm->mmap_sem);
1790                                 ret = 1;
1791                                 khugepaged_scan_shmem(mm, file->f_mapping,
1792                                                 pgoff, hpage);
1793                                 fput(file);
1794                         } else {
1795                                 ret = khugepaged_scan_pmd(mm, vma,
1796                                                 khugepaged_scan.address,
1797                                                 hpage);
1798                         }
1799                         /* move to next address */
1800                         khugepaged_scan.address += HPAGE_PMD_SIZE;
1801                         progress += HPAGE_PMD_NR;
1802                         if (ret)
1803                                 /* we released mmap_sem so break loop */
1804                                 goto breakouterloop_mmap_sem;
1805                         if (progress >= pages)
1806                                 goto breakouterloop;
1807                 }
1808         }
1809 breakouterloop:
1810         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
1811 breakouterloop_mmap_sem:
1812
1813         spin_lock(&khugepaged_mm_lock);
1814         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
1815         /*
1816          * Release the current mm_slot if this mm is about to die, or
1817          * if we scanned all vmas of this mm.
1818          */
1819         if (khugepaged_test_exit(mm) || !vma) {
1820                 /*
1821                  * Make sure that if mm_users is reaching zero while
1822                  * khugepaged runs here, khugepaged_exit will find
1823                  * mm_slot not pointing to the exiting mm.
1824                  */
1825                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
1826                         khugepaged_scan.mm_slot = list_entry(
1827                                 mm_slot->mm_node.next,
1828                                 struct mm_slot, mm_node);
1829                         khugepaged_scan.address = 0;
1830                 } else {
1831                         khugepaged_scan.mm_slot = NULL;
1832                         khugepaged_full_scans++;
1833                 }
1834
1835                 collect_mm_slot(mm_slot);
1836         }
1837
1838         return progress;
1839 }
1840
1841 static int khugepaged_has_work(void)
1842 {
1843         return !list_empty(&khugepaged_scan.mm_head) &&
1844                 khugepaged_enabled();
1845 }
1846
1847 static int khugepaged_wait_event(void)
1848 {
1849         return !list_empty(&khugepaged_scan.mm_head) ||
1850                 kthread_should_stop();
1851 }
1852
1853 static void khugepaged_do_scan(void)
1854 {
1855         struct page *hpage = NULL;
1856         unsigned int progress = 0, pass_through_head = 0;
1857         unsigned int pages = khugepaged_pages_to_scan;
1858         bool wait = true;
1859
1860         barrier(); /* write khugepaged_pages_to_scan to local stack */
1861
1862         while (progress < pages) {
1863                 if (!khugepaged_prealloc_page(&hpage, &wait))
1864                         break;
1865
1866                 cond_resched();
1867
1868                 if (unlikely(kthread_should_stop() || try_to_freeze()))
1869                         break;
1870
1871                 spin_lock(&khugepaged_mm_lock);
1872                 if (!khugepaged_scan.mm_slot)
1873                         pass_through_head++;
1874                 if (khugepaged_has_work() &&
1875                     pass_through_head < 2)
1876                         progress += khugepaged_scan_mm_slot(pages - progress,
1877                                                             &hpage);
1878                 else
1879                         progress = pages;
1880                 spin_unlock(&khugepaged_mm_lock);
1881         }
1882
1883         if (!IS_ERR_OR_NULL(hpage))
1884                 put_page(hpage);
1885 }
1886
1887 static bool khugepaged_should_wakeup(void)
1888 {
1889         return kthread_should_stop() ||
1890                time_after_eq(jiffies, khugepaged_sleep_expire);
1891 }
1892
1893 static void khugepaged_wait_work(void)
1894 {
1895         if (khugepaged_has_work()) {
1896                 const unsigned long scan_sleep_jiffies =
1897                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
1898
1899                 if (!scan_sleep_jiffies)
1900                         return;
1901
1902                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
1903                 wait_event_freezable_timeout(khugepaged_wait,
1904                                              khugepaged_should_wakeup(),
1905                                              scan_sleep_jiffies);
1906                 return;
1907         }
1908
1909         if (khugepaged_enabled())
1910                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
1911 }
1912
1913 static int khugepaged(void *none)
1914 {
1915         struct mm_slot *mm_slot;
1916
1917         set_freezable();
1918         set_user_nice(current, MAX_NICE);
1919
1920         while (!kthread_should_stop()) {
1921                 khugepaged_do_scan();
1922                 khugepaged_wait_work();
1923         }
1924
1925         spin_lock(&khugepaged_mm_lock);
1926         mm_slot = khugepaged_scan.mm_slot;
1927         khugepaged_scan.mm_slot = NULL;
1928         if (mm_slot)
1929                 collect_mm_slot(mm_slot);
1930         spin_unlock(&khugepaged_mm_lock);
1931         return 0;
1932 }
1933
1934 static void set_recommended_min_free_kbytes(void)
1935 {
1936         struct zone *zone;
1937         int nr_zones = 0;
1938         unsigned long recommended_min;
1939
1940         for_each_populated_zone(zone) {
1941                 /*
1942                  * We don't need to worry about fragmentation of
1943                  * ZONE_MOVABLE since it only has movable pages.
1944                  */
1945                 if (zone_idx(zone) > gfp_zone(GFP_USER))
1946                         continue;
1947
1948                 nr_zones++;
1949         }
1950
1951         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
1952         recommended_min = pageblock_nr_pages * nr_zones * 2;
1953
1954         /*
1955          * Make sure that on average at least two pageblocks are almost free
1956          * of another type, one for a migratetype to fall back to and a
1957          * second to avoid subsequent fallbacks of other types There are 3
1958          * MIGRATE_TYPES we care about.
1959          */
1960         recommended_min += pageblock_nr_pages * nr_zones *
1961                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
1962
1963         /* don't ever allow to reserve more than 5% of the lowmem */
1964         recommended_min = min(recommended_min,
1965                               (unsigned long) nr_free_buffer_pages() / 20);
1966         recommended_min <<= (PAGE_SHIFT-10);
1967
1968         if (recommended_min > min_free_kbytes) {
1969                 if (user_min_free_kbytes >= 0)
1970                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
1971                                 min_free_kbytes, recommended_min);
1972
1973                 min_free_kbytes = recommended_min;
1974         }
1975         setup_per_zone_wmarks();
1976 }
1977
1978 int start_stop_khugepaged(void)
1979 {
1980         int err = 0;
1981
1982         mutex_lock(&khugepaged_mutex);
1983         if (khugepaged_enabled()) {
1984                 if (!khugepaged_thread)
1985                         khugepaged_thread = kthread_run(khugepaged, NULL,
1986                                                         "khugepaged");
1987                 if (IS_ERR(khugepaged_thread)) {
1988                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
1989                         err = PTR_ERR(khugepaged_thread);
1990                         khugepaged_thread = NULL;
1991                         goto fail;
1992                 }
1993
1994                 if (!list_empty(&khugepaged_scan.mm_head))
1995                         wake_up_interruptible(&khugepaged_wait);
1996
1997                 set_recommended_min_free_kbytes();
1998         } else if (khugepaged_thread) {
1999                 kthread_stop(khugepaged_thread);
2000                 khugepaged_thread = NULL;
2001         }
2002 fail:
2003         mutex_unlock(&khugepaged_mutex);
2004         return err;
2005 }
2006
2007 void khugepaged_min_free_kbytes_update(void)
2008 {
2009         mutex_lock(&khugepaged_mutex);
2010         if (khugepaged_enabled() && khugepaged_thread)
2011                 set_recommended_min_free_kbytes();
2012         mutex_unlock(&khugepaged_mutex);
2013 }