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 int __khugepaged_enter(struct mm_struct *mm)
 417 {
 418         struct mm_slot *mm_slot;
 419         int wakeup;
 420
 421         mm_slot = alloc_mm_slot();
 422         if (!mm_slot)
 423                 return -ENOMEM;
 424
 425         /* __khugepaged_exit() must not run from under us */
 426         VM_BUG_ON_MM(atomic_read(&mm->mm_users) == 0, mm);
 427         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
 428                 free_mm_slot(mm_slot);
 429                 return 0;
 430         }
 431
 432         spin_lock(&khugepaged_mm_lock);
 433         insert_to_mm_slots_hash(mm, mm_slot);
 434         /*
 435          * Insert just behind the scanning cursor, to let the area settle
 436          * down a little.
 437          */
 438         wakeup = list_empty(&khugepaged_scan.mm_head);
 439         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
 440         spin_unlock(&khugepaged_mm_lock);
 441
 442         mmgrab(mm);
 443         if (wakeup)
 444                 wake_up_interruptible(&khugepaged_wait);
 445
 446         return 0;
 447 }
 448
 449 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
 450                                unsigned long vm_flags)
 451 {
 452         unsigned long hstart, hend;
 453         if (!vma->anon_vma)
 454                 /*
 455                  * Not yet faulted in so we will register later in the
 456                  * page fault if needed.
 457                  */
 458                 return 0;
 459         if (vma->vm_ops || (vm_flags & VM_NO_KHUGEPAGED))
 460                 /* khugepaged not yet working on file or special mappings */
 461                 return 0;
 462         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 463         hend = vma->vm_end & HPAGE_PMD_MASK;
 464         if (hstart < hend)
 465                 return khugepaged_enter(vma, vm_flags);
 466         return 0;
 467 }
 468
 469 void __khugepaged_exit(struct mm_struct *mm)
 470 {
 471         struct mm_slot *mm_slot;
 472         int free = 0;
 473
 474         spin_lock(&khugepaged_mm_lock);
 475         mm_slot = get_mm_slot(mm);
 476         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
 477                 hash_del(&mm_slot->hash);
 478                 list_del(&mm_slot->mm_node);
 479                 free = 1;
 480         }
 481         spin_unlock(&khugepaged_mm_lock);
 482
 483         if (free) {
 484                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
 485                 free_mm_slot(mm_slot);
 486                 mmdrop(mm);
 487         } else if (mm_slot) {
 488                 /*
 489                  * This is required to serialize against
 490                  * khugepaged_test_exit() (which is guaranteed to run
 491                  * under mmap sem read mode). Stop here (after we
 492                  * return all pagetables will be destroyed) until
 493                  * khugepaged has finished working on the pagetables
 494                  * under the mmap_sem.
 495                  */
 496                 down_write(&mm->mmap_sem);
 497                 up_write(&mm->mmap_sem);
 498         }
 499 }
 500
 501 static void release_pte_page(struct page *page)
 502 {
 503         dec_node_page_state(page, NR_ISOLATED_ANON + page_is_file_cache(page));
 504         unlock_page(page);
 505         putback_lru_page(page);
 506 }
 507
 508 static void release_pte_pages(pte_t *pte, pte_t *_pte)
 509 {
 510         while (--_pte >= pte) {
 511                 pte_t pteval = *_pte;
 512                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
 513                         release_pte_page(pte_page(pteval));
 514         }
 515 }
 516
 517 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
 518                                         unsigned long address,
 519                                         pte_t *pte)
 520 {
 521         struct page *page = NULL;
 522         pte_t *_pte;
 523         int none_or_zero = 0, result = 0, referenced = 0;
 524         bool writable = false;
 525
 526         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
 527              _pte++, address += PAGE_SIZE) {
 528                 pte_t pteval = *_pte;
 529                 if (pte_none(pteval) || (pte_present(pteval) &&
 530                                 is_zero_pfn(pte_pfn(pteval)))) {
 531                         if (!userfaultfd_armed(vma) &&
 532                             ++none_or_zero <= khugepaged_max_ptes_none) {
 533                                 continue;
 534                         } else {
 535                                 result = SCAN_EXCEED_NONE_PTE;
 536                                 goto out;
 537                         }
 538                 }
 539                 if (!pte_present(pteval)) {
 540                         result = SCAN_PTE_NON_PRESENT;
 541                         goto out;
 542                 }
 543                 page = vm_normal_page(vma, address, pteval);
 544                 if (unlikely(!page)) {
 545                         result = SCAN_PAGE_NULL;
 546                         goto out;
 547                 }
 548
 549                 /* TODO: teach khugepaged to collapse THP mapped with pte */
 550                 if (PageCompound(page)) {
 551                         result = SCAN_PAGE_COMPOUND;
 552                         goto out;
 553                 }
 554
 555                 VM_BUG_ON_PAGE(!PageAnon(page), page);
 556
 557                 /*
 558                  * We can do it before isolate_lru_page because the
 559                  * page can't be freed from under us. NOTE: PG_lock
 560                  * is needed to serialize against split_huge_page
 561                  * when invoked from the VM.
 562                  */
 563                 if (!trylock_page(page)) {
 564                         result = SCAN_PAGE_LOCK;
 565                         goto out;
 566                 }
 567
 568                 /*
 569                  * cannot use mapcount: can't collapse if there's a gup pin.
 570                  * The page must only be referenced by the scanned process
 571                  * and page swap cache.
 572                  */
 573                 if (page_count(page) != 1 + PageSwapCache(page)) {
 574                         unlock_page(page);
 575                         result = SCAN_PAGE_COUNT;
 576                         goto out;
 577                 }
 578                 if (pte_write(pteval)) {
 579                         writable = true;
 580                 } else {
 581                         if (PageSwapCache(page) &&
 582                             !reuse_swap_page(page, NULL)) {
 583                                 unlock_page(page);
 584                                 result = SCAN_SWAP_CACHE_PAGE;
 585                                 goto out;
 586                         }
 587                         /*
 588                          * Page is not in the swap cache. It can be collapsed
 589                          * into a THP.
 590                          */
 591                 }
 592
 593                 /*
 594                  * Isolate the page to avoid collapsing an hugepage
 595                  * currently in use by the VM.
 596                  */
 597                 if (isolate_lru_page(page)) {
 598                         unlock_page(page);
 599                         result = SCAN_DEL_PAGE_LRU;
 600                         goto out;
 601                 }
 602                 inc_node_page_state(page,
 603                                 NR_ISOLATED_ANON + page_is_file_cache(page));
 604                 VM_BUG_ON_PAGE(!PageLocked(page), page);
 605                 VM_BUG_ON_PAGE(PageLRU(page), page);
 606
 607                 /* There should be enough young pte to collapse the page */
 608                 if (pte_young(pteval) ||
 609                     page_is_young(page) || PageReferenced(page) ||
 610                     mmu_notifier_test_young(vma->vm_mm, address))
 611                         referenced++;
 612         }
 613
 614         if (unlikely(!writable)) {
 615                 result = SCAN_PAGE_RO;
 616         } else if (unlikely(!referenced)) {
 617                 result = SCAN_LACK_REFERENCED_PAGE;
 618         } else {
 619                 result = SCAN_SUCCEED;
 620                 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 621                                                     referenced, writable, result);
 622                 return 1;
 623         }
 624 out:
 625         release_pte_pages(pte, _pte);
 626         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 627                                             referenced, writable, result);
 628         return 0;
 629 }
 630
 631 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
 632                                       struct vm_area_struct *vma,
 633                                       unsigned long address,
 634                                       spinlock_t *ptl)
 635 {
 636         pte_t *_pte;
 637         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 638                                 _pte++, page++, address += PAGE_SIZE) {
 639                 pte_t pteval = *_pte;
 640                 struct page *src_page;
 641
 642                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
 643                         clear_user_highpage(page, address);
 644                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
 645                         if (is_zero_pfn(pte_pfn(pteval))) {
 646                                 /*
 647                                  * ptl mostly unnecessary.
 648                                  */
 649                                 spin_lock(ptl);
 650                                 /*
 651                                  * paravirt calls inside pte_clear here are
 652                                  * superfluous.
 653                                  */
 654                                 pte_clear(vma->vm_mm, address, _pte);
 655                                 spin_unlock(ptl);
 656                         }
 657                 } else {
 658                         src_page = pte_page(pteval);
 659                         copy_user_highpage(page, src_page, address, vma);
 660                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
 661                         release_pte_page(src_page);
 662                         /*
 663                          * ptl mostly unnecessary, but preempt has to
 664                          * be disabled to update the per-cpu stats
 665                          * inside page_remove_rmap().
 666                          */
 667                         spin_lock(ptl);
 668                         /*
 669                          * paravirt calls inside pte_clear here are
 670                          * superfluous.
 671                          */
 672                         pte_clear(vma->vm_mm, address, _pte);
 673                         page_remove_rmap(src_page, false);
 674                         spin_unlock(ptl);
 675                         free_page_and_swap_cache(src_page);
 676                 }
 677         }
 678 }
 679
 680 static void khugepaged_alloc_sleep(void)
 681 {
 682         DEFINE_WAIT(wait);
 683
 684         add_wait_queue(&khugepaged_wait, &wait);
 685         freezable_schedule_timeout_interruptible(
 686                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
 687         remove_wait_queue(&khugepaged_wait, &wait);
 688 }
 689
 690 static int khugepaged_node_load[MAX_NUMNODES];
 691
 692 static bool khugepaged_scan_abort(int nid)
 693 {
 694         int i;
 695
 696         /*
 697          * If node_reclaim_mode is disabled, then no extra effort is made to
 698          * allocate memory locally.
 699          */
 700         if (!node_reclaim_mode)
 701                 return false;
 702
 703         /* If there is a count for this node already, it must be acceptable */
 704         if (khugepaged_node_load[nid])
 705                 return false;
 706
 707         for (i = 0; i < MAX_NUMNODES; i++) {
 708                 if (!khugepaged_node_load[i])
 709                         continue;
 710                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
 711                         return true;
 712         }
 713         return false;
 714 }
 715
 716 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
 717 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
 718 {
 719         return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
 720 }
 721
 722 #ifdef CONFIG_NUMA
 723 static int khugepaged_find_target_node(void)
 724 {
 725         static int last_khugepaged_target_node = NUMA_NO_NODE;
 726         int nid, target_node = 0, max_value = 0;
 727
 728         /* find first node with max normal pages hit */
 729         for (nid = 0; nid < MAX_NUMNODES; nid++)
 730                 if (khugepaged_node_load[nid] > max_value) {
 731                         max_value = khugepaged_node_load[nid];
 732                         target_node = nid;
 733                 }
 734
 735         /* do some balance if several nodes have the same hit record */
 736         if (target_node <= last_khugepaged_target_node)
 737                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
 738                                 nid++)
 739                         if (max_value == khugepaged_node_load[nid]) {
 740                                 target_node = nid;
 741                                 break;
 742                         }
 743
 744         last_khugepaged_target_node = target_node;
 745         return target_node;
 746 }
 747
 748 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 749 {
 750         if (IS_ERR(*hpage)) {
 751                 if (!*wait)
 752                         return false;
 753
 754                 *wait = false;
 755                 *hpage = NULL;
 756                 khugepaged_alloc_sleep();
 757         } else if (*hpage) {
 758                 put_page(*hpage);
 759                 *hpage = NULL;
 760         }
 761
 762         return true;
 763 }
 764
 765 static struct page *
 766 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 767 {
 768         VM_BUG_ON_PAGE(*hpage, *hpage);
 769
 770         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
 771         if (unlikely(!*hpage)) {
 772                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 773                 *hpage = ERR_PTR(-ENOMEM);
 774                 return NULL;
 775         }
 776
 777         prep_transhuge_page(*hpage);
 778         count_vm_event(THP_COLLAPSE_ALLOC);
 779         return *hpage;
 780 }
 781 #else
 782 static int khugepaged_find_target_node(void)
 783 {
 784         return 0;
 785 }
 786
 787 static inline struct page *alloc_khugepaged_hugepage(void)
 788 {
 789         struct page *page;
 790
 791         page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
 792                            HPAGE_PMD_ORDER);
 793         if (page)
 794                 prep_transhuge_page(page);
 795         return page;
 796 }
 797
 798 static struct page *khugepaged_alloc_hugepage(bool *wait)
 799 {
 800         struct page *hpage;
 801
 802         do {
 803                 hpage = alloc_khugepaged_hugepage();
 804                 if (!hpage) {
 805                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 806                         if (!*wait)
 807                                 return NULL;
 808
 809                         *wait = false;
 810                         khugepaged_alloc_sleep();
 811                 } else
 812                         count_vm_event(THP_COLLAPSE_ALLOC);
 813         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
 814
 815         return hpage;
 816 }
 817
 818 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 819 {
 820         /*
 821          * If the hpage allocated earlier was briefly exposed in page cache
 822          * before collapse_file() failed, it is possible that racing lookups
 823          * have not yet completed, and would then be unpleasantly surprised by
 824          * finding the hpage reused for the same mapping at a different offset.
 825          * Just release the previous allocation if there is any danger of that.
 826          */
 827         if (*hpage && page_count(*hpage) > 1) {
 828                 put_page(*hpage);
 829                 *hpage = NULL;
 830         }
 831
 832         if (!*hpage)
 833                 *hpage = khugepaged_alloc_hugepage(wait);
 834
 835         if (unlikely(!*hpage))
 836                 return false;
 837
 838         return true;
 839 }
 840
 841 static struct page *
 842 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 843 {
 844         VM_BUG_ON(!*hpage);
 845
 846         return  *hpage;
 847 }
 848 #endif
 849
 850 static bool hugepage_vma_check(struct vm_area_struct *vma)
 851 {
 852         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
 853             (vma->vm_flags & VM_NOHUGEPAGE) ||
 854             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
 855                 return false;
 856         if (shmem_file(vma->vm_file)) {
 857                 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
 858                         return false;
 859                 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
 860                                 HPAGE_PMD_NR);
 861         }
 862         if (!vma->anon_vma || vma->vm_ops)
 863                 return false;
 864         if (is_vma_temporary_stack(vma))
 865                 return false;
 866         return !(vma->vm_flags & VM_NO_KHUGEPAGED);
 867 }
 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))
 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, ret = 0;
 912         struct vm_fault vmf = {
 913                 .vma = vma,
 914                 .address = address,
 915                 .flags = FAULT_FLAG_ALLOW_RETRY,
 916                 .pmd = pmd,
 917                 .pgoff = linear_page_index(vma, address),
 918         };
 919
 920         /* we only decide to swapin, if there is enough young ptes */
 921         if (referenced < HPAGE_PMD_NR/2) {
 922                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 923                 return false;
 924         }
 925         vmf.pte = pte_offset_map(pmd, address);
 926         for (; vmf.address < address + HPAGE_PMD_NR*PAGE_SIZE;
 927                         vmf.pte++, vmf.address += PAGE_SIZE) {
 928                 vmf.orig_pte = *vmf.pte;
 929                 if (!is_swap_pte(vmf.orig_pte))
 930                         continue;
 931                 swapped_in++;
 932                 ret = do_swap_page(&vmf);
 933
 934                 /* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */
 935                 if (ret & VM_FAULT_RETRY) {
 936                         down_read(&mm->mmap_sem);
 937                         if (hugepage_vma_revalidate(mm, address, &vmf.vma)) {
 938                                 /* vma is no longer available, don't continue to swapin */
 939                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 940                                 return false;
 941                         }
 942                         /* check if the pmd is still valid */
 943                         if (mm_find_pmd(mm, address) != pmd) {
 944                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 945                                 return false;
 946                         }
 947                 }
 948                 if (ret & VM_FAULT_ERROR) {
 949                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 950                         return false;
 951                 }
 952                 /* pte is unmapped now, we need to map it */
 953                 vmf.pte = pte_offset_map(pmd, vmf.address);
 954         }
 955         vmf.pte--;
 956         pte_unmap(vmf.pte);
 957         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
 958         return true;
 959 }
 960
 961 static void collapse_huge_page(struct mm_struct *mm,
 962                                    unsigned long address,
 963                                    struct page **hpage,
 964                                    int node, int referenced)
 965 {
 966         pmd_t *pmd, _pmd;
 967         pte_t *pte;
 968         pgtable_t pgtable;
 969         struct page *new_page;
 970         spinlock_t *pmd_ptl, *pte_ptl;
 971         int isolated = 0, result = 0;
 972         struct mem_cgroup *memcg;
 973         struct vm_area_struct *vma;
 974         unsigned long mmun_start;       /* For mmu_notifiers */
 975         unsigned long mmun_end;         /* For mmu_notifiers */
 976         gfp_t gfp;
 977
 978         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 979
 980         /* Only allocate from the target node */
 981         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
 982
 983         /*
 984          * Before allocating the hugepage, release the mmap_sem read lock.
 985          * The allocation can take potentially a long time if it involves
 986          * sync compaction, and we do not need to hold the mmap_sem during
 987          * that. We will recheck the vma after taking it again in write mode.
 988          */
 989         up_read(&mm->mmap_sem);
 990         new_page = khugepaged_alloc_page(hpage, gfp, node);
 991         if (!new_page) {
 992                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
 993                 goto out_nolock;
 994         }
 995
 996         /* Do not oom kill for khugepaged charges */
 997         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp | __GFP_NORETRY,
 998                                            &memcg, true))) {
 999                 result = SCAN_CGROUP_CHARGE_FAIL;
1000                 goto out_nolock;
1001         }
1002
1003         down_read(&mm->mmap_sem);
1004         result = hugepage_vma_revalidate(mm, address, &vma);
1005         if (result) {
1006                 mem_cgroup_cancel_charge(new_page, memcg, true);
1007                 up_read(&mm->mmap_sem);
1008                 goto out_nolock;
1009         }
1010
1011         pmd = mm_find_pmd(mm, address);
1012         if (!pmd) {
1013                 result = SCAN_PMD_NULL;
1014                 mem_cgroup_cancel_charge(new_page, memcg, true);
1015                 up_read(&mm->mmap_sem);
1016                 goto out_nolock;
1017         }
1018
1019         /*
1020          * __collapse_huge_page_swapin always returns with mmap_sem locked.
1021          * If it fails, we release mmap_sem and jump out_nolock.
1022          * Continuing to collapse causes inconsistency.
1023          */
1024         if (!__collapse_huge_page_swapin(mm, vma, address, pmd, referenced)) {
1025                 mem_cgroup_cancel_charge(new_page, memcg, true);
1026                 up_read(&mm->mmap_sem);
1027                 goto out_nolock;
1028         }
1029
1030         up_read(&mm->mmap_sem);
1031         /*
1032          * Prevent all access to pagetables with the exception of
1033          * gup_fast later handled by the ptep_clear_flush and the VM
1034          * handled by the anon_vma lock + PG_lock.
1035          */
1036         down_write(&mm->mmap_sem);
1037         result = hugepage_vma_revalidate(mm, address, &vma);
1038         if (result)
1039                 goto out;
1040         /* check if the pmd is still valid */
1041         if (mm_find_pmd(mm, address) != pmd)
1042                 goto out;
1043
1044         anon_vma_lock_write(vma->anon_vma);
1045
1046         pte = pte_offset_map(pmd, address);
1047         pte_ptl = pte_lockptr(mm, pmd);
1048
1049         mmun_start = address;
1050         mmun_end   = address + HPAGE_PMD_SIZE;
1051         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1052         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1053         /*
1054          * After this gup_fast can't run anymore. This also removes
1055          * any huge TLB entry from the CPU so we won't allow
1056          * huge and small TLB entries for the same virtual address
1057          * to avoid the risk of CPU bugs in that area.
1058          */
1059         _pmd = pmdp_collapse_flush(vma, address, pmd);
1060         spin_unlock(pmd_ptl);
1061         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1062         tlb_remove_table_sync_one();
1063
1064         spin_lock(pte_ptl);
1065         isolated = __collapse_huge_page_isolate(vma, address, pte);
1066         spin_unlock(pte_ptl);
1067
1068         if (unlikely(!isolated)) {
1069                 pte_unmap(pte);
1070                 spin_lock(pmd_ptl);
1071                 BUG_ON(!pmd_none(*pmd));
1072                 /*
1073                  * We can only use set_pmd_at when establishing
1074                  * hugepmds and never for establishing regular pmds that
1075                  * points to regular pagetables. Use pmd_populate for that
1076                  */
1077                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1078                 spin_unlock(pmd_ptl);
1079                 anon_vma_unlock_write(vma->anon_vma);
1080                 result = SCAN_FAIL;
1081                 goto out;
1082         }
1083
1084         /*
1085          * All pages are isolated and locked so anon_vma rmap
1086          * can't run anymore.
1087          */
1088         anon_vma_unlock_write(vma->anon_vma);
1089
1090         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
1091         pte_unmap(pte);
1092         __SetPageUptodate(new_page);
1093         pgtable = pmd_pgtable(_pmd);
1094
1095         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1096         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1097
1098         /*
1099          * spin_lock() below is not the equivalent of smp_wmb(), so
1100          * this is needed to avoid the copy_huge_page writes to become
1101          * visible after the set_pmd_at() write.
1102          */
1103         smp_wmb();
1104
1105         spin_lock(pmd_ptl);
1106         BUG_ON(!pmd_none(*pmd));
1107         page_add_new_anon_rmap(new_page, vma, address, true);
1108         mem_cgroup_commit_charge(new_page, memcg, false, true);
1109         lru_cache_add_active_or_unevictable(new_page, vma);
1110         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1111         set_pmd_at(mm, address, pmd, _pmd);
1112         update_mmu_cache_pmd(vma, address, pmd);
1113         spin_unlock(pmd_ptl);
1114
1115         *hpage = NULL;
1116
1117         khugepaged_pages_collapsed++;
1118         result = SCAN_SUCCEED;
1119 out_up_write:
1120         up_write(&mm->mmap_sem);
1121 out_nolock:
1122         trace_mm_collapse_huge_page(mm, isolated, result);
1123         return;
1124 out:
1125         mem_cgroup_cancel_charge(new_page, memcg, true);
1126         goto out_up_write;
1127 }
1128
1129 static int khugepaged_scan_pmd(struct mm_struct *mm,
1130                                struct vm_area_struct *vma,
1131                                unsigned long address,
1132                                struct page **hpage)
1133 {
1134         pmd_t *pmd;
1135         pte_t *pte, *_pte;
1136         int ret = 0, none_or_zero = 0, result = 0, referenced = 0;
1137         struct page *page = NULL;
1138         unsigned long _address;
1139         spinlock_t *ptl;
1140         int node = NUMA_NO_NODE, unmapped = 0;
1141         bool writable = false;
1142
1143         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1144
1145         pmd = mm_find_pmd(mm, address);
1146         if (!pmd) {
1147                 result = SCAN_PMD_NULL;
1148                 goto out;
1149         }
1150
1151         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1152         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1153         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1154              _pte++, _address += PAGE_SIZE) {
1155                 pte_t pteval = *_pte;
1156                 if (is_swap_pte(pteval)) {
1157                         if (++unmapped <= khugepaged_max_ptes_swap) {
1158                                 continue;
1159                         } else {
1160                                 result = SCAN_EXCEED_SWAP_PTE;
1161                                 goto out_unmap;
1162                         }
1163                 }
1164                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1165                         if (!userfaultfd_armed(vma) &&
1166                             ++none_or_zero <= khugepaged_max_ptes_none) {
1167                                 continue;
1168                         } else {
1169                                 result = SCAN_EXCEED_NONE_PTE;
1170                                 goto out_unmap;
1171                         }
1172                 }
1173                 if (!pte_present(pteval)) {
1174                         result = SCAN_PTE_NON_PRESENT;
1175                         goto out_unmap;
1176                 }
1177                 if (pte_write(pteval))
1178                         writable = true;
1179
1180                 page = vm_normal_page(vma, _address, pteval);
1181                 if (unlikely(!page)) {
1182                         result = SCAN_PAGE_NULL;
1183                         goto out_unmap;
1184                 }
1185
1186                 /* TODO: teach khugepaged to collapse THP mapped with pte */
1187                 if (PageCompound(page)) {
1188                         result = SCAN_PAGE_COMPOUND;
1189                         goto out_unmap;
1190                 }
1191
1192                 /*
1193                  * Record which node the original page is from and save this
1194                  * information to khugepaged_node_load[].
1195                  * Khupaged will allocate hugepage from the node has the max
1196                  * hit record.
1197                  */
1198                 node = page_to_nid(page);
1199                 if (khugepaged_scan_abort(node)) {
1200                         result = SCAN_SCAN_ABORT;
1201                         goto out_unmap;
1202                 }
1203                 khugepaged_node_load[node]++;
1204                 if (!PageLRU(page)) {
1205                         result = SCAN_PAGE_LRU;
1206                         goto out_unmap;
1207                 }
1208                 if (PageLocked(page)) {
1209                         result = SCAN_PAGE_LOCK;
1210                         goto out_unmap;
1211                 }
1212                 if (!PageAnon(page)) {
1213                         result = SCAN_PAGE_ANON;
1214                         goto out_unmap;
1215                 }
1216
1217                 /*
1218                  * cannot use mapcount: can't collapse if there's a gup pin.
1219                  * The page must only be referenced by the scanned process
1220                  * and page swap cache.
1221                  */
1222                 if (page_count(page) != 1 + PageSwapCache(page)) {
1223                         result = SCAN_PAGE_COUNT;
1224                         goto out_unmap;
1225                 }
1226                 if (pte_young(pteval) ||
1227                     page_is_young(page) || PageReferenced(page) ||
1228                     mmu_notifier_test_young(vma->vm_mm, address))
1229                         referenced++;
1230         }
1231         if (writable) {
1232                 if (referenced) {
1233                         result = SCAN_SUCCEED;
1234                         ret = 1;
1235                 } else {
1236                         result = SCAN_LACK_REFERENCED_PAGE;
1237                 }
1238         } else {
1239                 result = SCAN_PAGE_RO;
1240         }
1241 out_unmap:
1242         pte_unmap_unlock(pte, ptl);
1243         if (ret) {
1244                 node = khugepaged_find_target_node();
1245                 /* collapse_huge_page will return with the mmap_sem released */
1246                 collapse_huge_page(mm, address, hpage, node, referenced);
1247         }
1248 out:
1249         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1250                                      none_or_zero, result, unmapped);
1251         return ret;
1252 }
1253
1254 static void collect_mm_slot(struct mm_slot *mm_slot)
1255 {
1256         struct mm_struct *mm = mm_slot->mm;
1257
1258         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
1259
1260         if (khugepaged_test_exit(mm)) {
1261                 /* free mm_slot */
1262                 hash_del(&mm_slot->hash);
1263                 list_del(&mm_slot->mm_node);
1264
1265                 /*
1266                  * Not strictly needed because the mm exited already.
1267                  *
1268                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1269                  */
1270
1271                 /* khugepaged_mm_lock actually not necessary for the below */
1272                 free_mm_slot(mm_slot);
1273                 mmdrop(mm);
1274         }
1275 }
1276
1277 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
1278 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1279 {
1280         struct vm_area_struct *vma;
1281         struct mm_struct *mm;
1282         unsigned long addr;
1283         pmd_t *pmd, _pmd;
1284
1285         i_mmap_lock_write(mapping);
1286         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1287                 /* probably overkill */
1288                 if (vma->anon_vma)
1289                         continue;
1290                 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1291                 if (addr & ~HPAGE_PMD_MASK)
1292                         continue;
1293                 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1294                         continue;
1295                 mm = vma->vm_mm;
1296                 pmd = mm_find_pmd(mm, addr);
1297                 if (!pmd)
1298                         continue;
1299                 /*
1300                  * We need exclusive mmap_sem to retract page table.
1301                  * If trylock fails we would end up with pte-mapped THP after
1302                  * re-fault. Not ideal, but it's more important to not disturb
1303                  * the system too much.
1304                  */
1305                 if (down_write_trylock(&mm->mmap_sem)) {
1306                         if (!khugepaged_test_exit(mm)) {
1307                                 spinlock_t *ptl;
1308                                 unsigned long end = addr + HPAGE_PMD_SIZE;
1309
1310                                 mmu_notifier_invalidate_range_start(mm, addr,
1311                                                                     end);
1312                                 ptl = pmd_lock(mm, pmd);
1313                                 /* assume page table is clear */
1314                                 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1315                                 spin_unlock(ptl);
1316                                 atomic_long_dec(&mm->nr_ptes);
1317                                 tlb_remove_table_sync_one();
1318                                 pte_free(mm, pmd_pgtable(_pmd));
1319                                 mmu_notifier_invalidate_range_end(mm, addr,
1320                                                                   end);
1321                         }
1322                         up_write(&mm->mmap_sem);
1323                 }
1324         }
1325         i_mmap_unlock_write(mapping);
1326 }
1327
1328 /**
1329  * collapse_shmem - collapse small tmpfs/shmem pages into huge one.
1330  *
1331  * Basic scheme is simple, details are more complex:
1332  *  - allocate and lock a new huge page;
1333  *  - scan over radix tree replacing old pages the new one
1334  *    + swap in pages if necessary;
1335  *    + fill in gaps;
1336  *    + keep old pages around in case if rollback is required;
1337  *  - if replacing succeed:
1338  *    + copy data over;
1339  *    + free old pages;
1340  *    + unlock huge page;
1341  *  - if replacing failed;
1342  *    + put all pages back and unfreeze them;
1343  *    + restore gaps in the radix-tree;
1344  *    + unlock and free huge page;
1345  */
1346 static void collapse_shmem(struct mm_struct *mm,
1347                 struct address_space *mapping, pgoff_t start,
1348                 struct page **hpage, int node)
1349 {
1350         gfp_t gfp;
1351         struct page *page, *new_page, *tmp;
1352         struct mem_cgroup *memcg;
1353         pgoff_t index, end = start + HPAGE_PMD_NR;
1354         LIST_HEAD(pagelist);
1355         struct radix_tree_iter iter;
1356         void **slot;
1357         int nr_none = 0, result = SCAN_SUCCEED;
1358
1359         VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1360
1361         /* Only allocate from the target node */
1362         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1363
1364         new_page = khugepaged_alloc_page(hpage, gfp, node);
1365         if (!new_page) {
1366                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1367                 goto out;
1368         }
1369
1370         /* Do not oom kill for khugepaged charges */
1371         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp | __GFP_NORETRY,
1372                                            &memcg, true))) {
1373                 result = SCAN_CGROUP_CHARGE_FAIL;
1374                 goto out;
1375         }
1376
1377         __SetPageLocked(new_page);
1378         __SetPageSwapBacked(new_page);
1379         new_page->index = start;
1380         new_page->mapping = mapping;
1381
1382         /*
1383          * At this point the new_page is locked and not up-to-date.
1384          * It's safe to insert it into the page cache, because nobody would
1385          * be able to map it or use it in another way until we unlock it.
1386          */
1387
1388         index = start;
1389         spin_lock_irq(&mapping->tree_lock);
1390         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1391                 int n = min(iter.index, end) - index;
1392
1393                 /*
1394                  * Stop if extent has been hole-punched, and is now completely
1395                  * empty (the more obvious i_size_read() check would take an
1396                  * irq-unsafe seqlock on 32-bit).
1397                  */
1398                 if (n >= HPAGE_PMD_NR) {
1399                         result = SCAN_TRUNCATED;
1400                         goto tree_locked;
1401                 }
1402
1403                 /*
1404                  * Handle holes in the radix tree: charge it from shmem and
1405                  * insert relevant subpage of new_page into the radix-tree.
1406                  */
1407                 if (n && !shmem_charge(mapping->host, n)) {
1408                         result = SCAN_FAIL;
1409                         goto tree_locked;
1410                 }
1411                 for (; index < min(iter.index, end); index++) {
1412                         radix_tree_insert(&mapping->page_tree, index,
1413                                         new_page + (index % HPAGE_PMD_NR));
1414                 }
1415                 nr_none += n;
1416
1417                 /* We are done. */
1418                 if (index >= end)
1419                         break;
1420
1421                 page = radix_tree_deref_slot_protected(slot,
1422                                 &mapping->tree_lock);
1423                 if (radix_tree_exceptional_entry(page) || !PageUptodate(page)) {
1424                         spin_unlock_irq(&mapping->tree_lock);
1425                         /* swap in or instantiate fallocated page */
1426                         if (shmem_getpage(mapping->host, index, &page,
1427                                                 SGP_NOHUGE)) {
1428                                 result = SCAN_FAIL;
1429                                 goto tree_unlocked;
1430                         }
1431                 } else if (trylock_page(page)) {
1432                         get_page(page);
1433                         spin_unlock_irq(&mapping->tree_lock);
1434                 } else {
1435                         result = SCAN_PAGE_LOCK;
1436                         goto tree_locked;
1437                 }
1438
1439                 /*
1440                  * The page must be locked, so we can drop the tree_lock
1441                  * without racing with truncate.
1442                  */
1443                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1444                 VM_BUG_ON_PAGE(!PageUptodate(page), page);
1445
1446                 /*
1447                  * If file was truncated then extended, or hole-punched, before
1448                  * we locked the first page, then a THP might be there already.
1449                  */
1450                 if (PageTransCompound(page)) {
1451                         result = SCAN_PAGE_COMPOUND;
1452                         goto out_unlock;
1453                 }
1454
1455                 if (page_mapping(page) != mapping) {
1456                         result = SCAN_TRUNCATED;
1457                         goto out_unlock;
1458                 }
1459
1460                 if (isolate_lru_page(page)) {
1461                         result = SCAN_DEL_PAGE_LRU;
1462                         goto out_unlock;
1463                 }
1464
1465                 if (page_mapped(page))
1466                         unmap_mapping_range(mapping, index << PAGE_SHIFT,
1467                                         PAGE_SIZE, 0);
1468
1469                 spin_lock_irq(&mapping->tree_lock);
1470
1471                 slot = radix_tree_lookup_slot(&mapping->page_tree, index);
1472                 VM_BUG_ON_PAGE(page != radix_tree_deref_slot_protected(slot,
1473                                         &mapping->tree_lock), page);
1474                 VM_BUG_ON_PAGE(page_mapped(page), page);
1475
1476                 /*
1477                  * The page is expected to have page_count() == 3:
1478                  *  - we hold a pin on it;
1479                  *  - one reference from radix tree;
1480                  *  - one from isolate_lru_page;
1481                  */
1482                 if (!page_ref_freeze(page, 3)) {
1483                         result = SCAN_PAGE_COUNT;
1484                         spin_unlock_irq(&mapping->tree_lock);
1485                         putback_lru_page(page);
1486                         goto out_unlock;
1487                 }
1488
1489                 /*
1490                  * Add the page to the list to be able to undo the collapse if
1491                  * something go wrong.
1492                  */
1493                 list_add_tail(&page->lru, &pagelist);
1494
1495                 /* Finally, replace with the new page. */
1496                 radix_tree_replace_slot(&mapping->page_tree, slot,
1497                                 new_page + (index % HPAGE_PMD_NR));
1498
1499                 slot = radix_tree_iter_resume(slot, &iter);
1500                 index++;
1501                 continue;
1502 out_unlock:
1503                 unlock_page(page);
1504                 put_page(page);
1505                 goto tree_unlocked;
1506         }
1507
1508         /*
1509          * Handle hole in radix tree at the end of the range.
1510          * This code only triggers if there's nothing in radix tree
1511          * beyond 'end'.
1512          */
1513         if (index < end) {
1514                 int n = end - index;
1515
1516                 /* Stop if extent has been truncated, and is now empty */
1517                 if (n >= HPAGE_PMD_NR) {
1518                         result = SCAN_TRUNCATED;
1519                         goto tree_locked;
1520                 }
1521                 if (!shmem_charge(mapping->host, n)) {
1522                         result = SCAN_FAIL;
1523                         goto tree_locked;
1524                 }
1525                 for (; index < end; index++) {
1526                         radix_tree_insert(&mapping->page_tree, index,
1527                                         new_page + (index % HPAGE_PMD_NR));
1528                 }
1529                 nr_none += n;
1530         }
1531
1532         __inc_node_page_state(new_page, NR_SHMEM_THPS);
1533         if (nr_none) {
1534                 struct zone *zone = page_zone(new_page);
1535
1536                 __mod_node_page_state(zone->zone_pgdat, NR_FILE_PAGES, nr_none);
1537                 __mod_node_page_state(zone->zone_pgdat, NR_SHMEM, nr_none);
1538         }
1539
1540 tree_locked:
1541         spin_unlock_irq(&mapping->tree_lock);
1542 tree_unlocked:
1543
1544         if (result == SCAN_SUCCEED) {
1545                 /*
1546                  * Replacing old pages with new one has succeed, now we need to
1547                  * copy the content and free old pages.
1548                  */
1549                 index = start;
1550                 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1551                         while (index < page->index) {
1552                                 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1553                                 index++;
1554                         }
1555                         copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1556                                         page);
1557                         list_del(&page->lru);
1558                         page->mapping = NULL;
1559                         page_ref_unfreeze(page, 1);
1560                         ClearPageActive(page);
1561                         ClearPageUnevictable(page);
1562                         unlock_page(page);
1563                         put_page(page);
1564                         index++;
1565                 }
1566                 while (index < end) {
1567                         clear_highpage(new_page + (index % HPAGE_PMD_NR));
1568                         index++;
1569                 }
1570
1571                 SetPageUptodate(new_page);
1572                 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1573                 set_page_dirty(new_page);
1574                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1575                 lru_cache_add_anon(new_page);
1576
1577                 /*
1578                  * Remove pte page tables, so we can re-fault the page as huge.
1579                  */
1580                 retract_page_tables(mapping, start);
1581                 *hpage = NULL;
1582         } else {
1583                 /* Something went wrong: rollback changes to the radix-tree */
1584                 spin_lock_irq(&mapping->tree_lock);
1585                 mapping->nrpages -= nr_none;
1586                 shmem_uncharge(mapping->host, nr_none);
1587
1588                 radix_tree_for_each_slot(slot, &mapping->page_tree, &iter,
1589                                 start) {
1590                         if (iter.index >= end)
1591                                 break;
1592                         page = list_first_entry_or_null(&pagelist,
1593                                         struct page, lru);
1594                         if (!page || iter.index < page->index) {
1595                                 if (!nr_none)
1596                                         break;
1597                                 nr_none--;
1598                                 /* Put holes back where they were */
1599                                 radix_tree_delete(&mapping->page_tree,
1600                                                   iter.index);
1601                                 continue;
1602                         }
1603
1604                         VM_BUG_ON_PAGE(page->index != iter.index, page);
1605
1606                         /* Unfreeze the page. */
1607                         list_del(&page->lru);
1608                         page_ref_unfreeze(page, 2);
1609                         radix_tree_replace_slot(&mapping->page_tree,
1610                                                 slot, page);
1611                         slot = radix_tree_iter_resume(slot, &iter);
1612                         spin_unlock_irq(&mapping->tree_lock);
1613                         unlock_page(page);
1614                         putback_lru_page(page);
1615                         spin_lock_irq(&mapping->tree_lock);
1616                 }
1617                 VM_BUG_ON(nr_none);
1618                 spin_unlock_irq(&mapping->tree_lock);
1619
1620                 mem_cgroup_cancel_charge(new_page, memcg, true);
1621                 new_page->mapping = NULL;
1622         }
1623
1624         unlock_page(new_page);
1625 out:
1626         VM_BUG_ON(!list_empty(&pagelist));
1627         /* TODO: tracepoints */
1628 }
1629
1630 static void khugepaged_scan_shmem(struct mm_struct *mm,
1631                 struct address_space *mapping,
1632                 pgoff_t start, struct page **hpage)
1633 {
1634         struct page *page = NULL;
1635         struct radix_tree_iter iter;
1636         void **slot;
1637         int present, swap;
1638         int node = NUMA_NO_NODE;
1639         int result = SCAN_SUCCEED;
1640
1641         present = 0;
1642         swap = 0;
1643         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1644         rcu_read_lock();
1645         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1646                 if (iter.index >= start + HPAGE_PMD_NR)
1647                         break;
1648
1649                 page = radix_tree_deref_slot(slot);
1650                 if (radix_tree_deref_retry(page)) {
1651                         slot = radix_tree_iter_retry(&iter);
1652                         continue;
1653                 }
1654
1655                 if (radix_tree_exception(page)) {
1656                         if (++swap > khugepaged_max_ptes_swap) {
1657                                 result = SCAN_EXCEED_SWAP_PTE;
1658                                 break;
1659                         }
1660                         continue;
1661                 }
1662
1663                 if (PageTransCompound(page)) {
1664                         result = SCAN_PAGE_COMPOUND;
1665                         break;
1666                 }
1667
1668                 node = page_to_nid(page);
1669                 if (khugepaged_scan_abort(node)) {
1670                         result = SCAN_SCAN_ABORT;
1671                         break;
1672                 }
1673                 khugepaged_node_load[node]++;
1674
1675                 if (!PageLRU(page)) {
1676                         result = SCAN_PAGE_LRU;
1677                         break;
1678                 }
1679
1680                 if (page_count(page) != 1 + page_mapcount(page)) {
1681                         result = SCAN_PAGE_COUNT;
1682                         break;
1683                 }
1684
1685                 /*
1686                  * We probably should check if the page is referenced here, but
1687                  * nobody would transfer pte_young() to PageReferenced() for us.
1688                  * And rmap walk here is just too costly...
1689                  */
1690
1691                 present++;
1692
1693                 if (need_resched()) {
1694                         slot = radix_tree_iter_resume(slot, &iter);
1695                         cond_resched_rcu();
1696                 }
1697         }
1698         rcu_read_unlock();
1699
1700         if (result == SCAN_SUCCEED) {
1701                 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
1702                         result = SCAN_EXCEED_NONE_PTE;
1703                 } else {
1704                         node = khugepaged_find_target_node();
1705                         collapse_shmem(mm, mapping, start, hpage, node);
1706                 }
1707         }
1708
1709         /* TODO: tracepoints */
1710 }
1711 #else
1712 static void khugepaged_scan_shmem(struct mm_struct *mm,
1713                 struct address_space *mapping,
1714                 pgoff_t start, struct page **hpage)
1715 {
1716         BUILD_BUG();
1717 }
1718 #endif
1719
1720 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
1721                                             struct page **hpage)
1722         __releases(&khugepaged_mm_lock)
1723         __acquires(&khugepaged_mm_lock)
1724 {
1725         struct mm_slot *mm_slot;
1726         struct mm_struct *mm;
1727         struct vm_area_struct *vma;
1728         int progress = 0;
1729
1730         VM_BUG_ON(!pages);
1731         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
1732
1733         if (khugepaged_scan.mm_slot)
1734                 mm_slot = khugepaged_scan.mm_slot;
1735         else {
1736                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
1737                                      struct mm_slot, mm_node);
1738                 khugepaged_scan.address = 0;
1739                 khugepaged_scan.mm_slot = mm_slot;
1740         }
1741         spin_unlock(&khugepaged_mm_lock);
1742
1743         mm = mm_slot->mm;
1744         /*
1745          * Don't wait for semaphore (to avoid long wait times).  Just move to
1746          * the next mm on the list.
1747          */
1748         vma = NULL;
1749         if (unlikely(!down_read_trylock(&mm->mmap_sem)))
1750                 goto breakouterloop_mmap_sem;
1751         if (likely(!khugepaged_test_exit(mm)))
1752                 vma = find_vma(mm, khugepaged_scan.address);
1753
1754         progress++;
1755         for (; vma; vma = vma->vm_next) {
1756                 unsigned long hstart, hend;
1757
1758                 cond_resched();
1759                 if (unlikely(khugepaged_test_exit(mm))) {
1760                         progress++;
1761                         break;
1762                 }
1763                 if (!hugepage_vma_check(vma)) {
1764 skip:
1765                         progress++;
1766                         continue;
1767                 }
1768                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1769                 hend = vma->vm_end & HPAGE_PMD_MASK;
1770                 if (hstart >= hend)
1771                         goto skip;
1772                 if (khugepaged_scan.address > hend)
1773                         goto skip;
1774                 if (khugepaged_scan.address < hstart)
1775                         khugepaged_scan.address = hstart;
1776                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
1777
1778                 while (khugepaged_scan.address < hend) {
1779                         int ret;
1780                         cond_resched();
1781                         if (unlikely(khugepaged_test_exit(mm)))
1782                                 goto breakouterloop;
1783
1784                         VM_BUG_ON(khugepaged_scan.address < hstart ||
1785                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
1786                                   hend);
1787                         if (shmem_file(vma->vm_file)) {
1788                                 struct file *file;
1789                                 pgoff_t pgoff = linear_page_index(vma,
1790                                                 khugepaged_scan.address);
1791                                 if (!shmem_huge_enabled(vma))
1792                                         goto skip;
1793                                 file = get_file(vma->vm_file);
1794                                 up_read(&mm->mmap_sem);
1795                                 ret = 1;
1796                                 khugepaged_scan_shmem(mm, file->f_mapping,
1797                                                 pgoff, hpage);
1798                                 fput(file);
1799                         } else {
1800                                 ret = khugepaged_scan_pmd(mm, vma,
1801                                                 khugepaged_scan.address,
1802                                                 hpage);
1803                         }
1804                         /* move to next address */
1805                         khugepaged_scan.address += HPAGE_PMD_SIZE;
1806                         progress += HPAGE_PMD_NR;
1807                         if (ret)
1808                                 /* we released mmap_sem so break loop */
1809                                 goto breakouterloop_mmap_sem;
1810                         if (progress >= pages)
1811                                 goto breakouterloop;
1812                 }
1813         }
1814 breakouterloop:
1815         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
1816 breakouterloop_mmap_sem:
1817
1818         spin_lock(&khugepaged_mm_lock);
1819         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
1820         /*
1821          * Release the current mm_slot if this mm is about to die, or
1822          * if we scanned all vmas of this mm.
1823          */
1824         if (khugepaged_test_exit(mm) || !vma) {
1825                 /*
1826                  * Make sure that if mm_users is reaching zero while
1827                  * khugepaged runs here, khugepaged_exit will find
1828                  * mm_slot not pointing to the exiting mm.
1829                  */
1830                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
1831                         khugepaged_scan.mm_slot = list_entry(
1832                                 mm_slot->mm_node.next,
1833                                 struct mm_slot, mm_node);
1834                         khugepaged_scan.address = 0;
1835                 } else {
1836                         khugepaged_scan.mm_slot = NULL;
1837                         khugepaged_full_scans++;
1838                 }
1839
1840                 collect_mm_slot(mm_slot);
1841         }
1842
1843         return progress;
1844 }
1845
1846 static int khugepaged_has_work(void)
1847 {
1848         return !list_empty(&khugepaged_scan.mm_head) &&
1849                 khugepaged_enabled();
1850 }
1851
1852 static int khugepaged_wait_event(void)
1853 {
1854         return !list_empty(&khugepaged_scan.mm_head) ||
1855                 kthread_should_stop();
1856 }
1857
1858 static void khugepaged_do_scan(void)
1859 {
1860         struct page *hpage = NULL;
1861         unsigned int progress = 0, pass_through_head = 0;
1862         unsigned int pages = khugepaged_pages_to_scan;
1863         bool wait = true;
1864
1865         barrier(); /* write khugepaged_pages_to_scan to local stack */
1866
1867         while (progress < pages) {
1868                 if (!khugepaged_prealloc_page(&hpage, &wait))
1869                         break;
1870
1871                 cond_resched();
1872
1873                 if (unlikely(kthread_should_stop() || try_to_freeze()))
1874                         break;
1875
1876                 spin_lock(&khugepaged_mm_lock);
1877                 if (!khugepaged_scan.mm_slot)
1878                         pass_through_head++;
1879                 if (khugepaged_has_work() &&
1880                     pass_through_head < 2)
1881                         progress += khugepaged_scan_mm_slot(pages - progress,
1882                                                             &hpage);
1883                 else
1884                         progress = pages;
1885                 spin_unlock(&khugepaged_mm_lock);
1886         }
1887
1888         if (!IS_ERR_OR_NULL(hpage))
1889                 put_page(hpage);
1890 }
1891
1892 static bool khugepaged_should_wakeup(void)
1893 {
1894         return kthread_should_stop() ||
1895                time_after_eq(jiffies, khugepaged_sleep_expire);
1896 }
1897
1898 static void khugepaged_wait_work(void)
1899 {
1900         if (khugepaged_has_work()) {
1901                 const unsigned long scan_sleep_jiffies =
1902                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
1903
1904                 if (!scan_sleep_jiffies)
1905                         return;
1906
1907                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
1908                 wait_event_freezable_timeout(khugepaged_wait,
1909                                              khugepaged_should_wakeup(),
1910                                              scan_sleep_jiffies);
1911                 return;
1912         }
1913
1914         if (khugepaged_enabled())
1915                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
1916 }
1917
1918 static int khugepaged(void *none)
1919 {
1920         struct mm_slot *mm_slot;
1921
1922         set_freezable();
1923         set_user_nice(current, MAX_NICE);
1924
1925         while (!kthread_should_stop()) {
1926                 khugepaged_do_scan();
1927                 khugepaged_wait_work();
1928         }
1929
1930         spin_lock(&khugepaged_mm_lock);
1931         mm_slot = khugepaged_scan.mm_slot;
1932         khugepaged_scan.mm_slot = NULL;
1933         if (mm_slot)
1934                 collect_mm_slot(mm_slot);
1935         spin_unlock(&khugepaged_mm_lock);
1936         return 0;
1937 }
1938
1939 static void set_recommended_min_free_kbytes(void)
1940 {
1941         struct zone *zone;
1942         int nr_zones = 0;
1943         unsigned long recommended_min;
1944
1945         for_each_populated_zone(zone)
1946                 nr_zones++;
1947
1948         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
1949         recommended_min = pageblock_nr_pages * nr_zones * 2;
1950
1951         /*
1952          * Make sure that on average at least two pageblocks are almost free
1953          * of another type, one for a migratetype to fall back to and a
1954          * second to avoid subsequent fallbacks of other types There are 3
1955          * MIGRATE_TYPES we care about.
1956          */
1957         recommended_min += pageblock_nr_pages * nr_zones *
1958                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
1959
1960         /* don't ever allow to reserve more than 5% of the lowmem */
1961         recommended_min = min(recommended_min,
1962                               (unsigned long) nr_free_buffer_pages() / 20);
1963         recommended_min <<= (PAGE_SHIFT-10);
1964
1965         if (recommended_min > min_free_kbytes) {
1966                 if (user_min_free_kbytes >= 0)
1967                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
1968                                 min_free_kbytes, recommended_min);
1969
1970                 min_free_kbytes = recommended_min;
1971         }
1972         setup_per_zone_wmarks();
1973 }
1974
1975 int start_stop_khugepaged(void)
1976 {
1977         int err = 0;
1978
1979         mutex_lock(&khugepaged_mutex);
1980         if (khugepaged_enabled()) {
1981                 if (!khugepaged_thread)
1982                         khugepaged_thread = kthread_run(khugepaged, NULL,
1983                                                         "khugepaged");
1984                 if (IS_ERR(khugepaged_thread)) {
1985                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
1986                         err = PTR_ERR(khugepaged_thread);
1987                         khugepaged_thread = NULL;
1988                         goto fail;
1989                 }
1990
1991                 if (!list_empty(&khugepaged_scan.mm_head))
1992                         wake_up_interruptible(&khugepaged_wait);
1993
1994                 set_recommended_min_free_kbytes();
1995         } else if (khugepaged_thread) {
1996                 kthread_stop(khugepaged_thread);
1997                 khugepaged_thread = NULL;
1998         }
1999 fail:
2000         mutex_unlock(&khugepaged_mutex);
2001         return err;
2002 }
2003
2004 void khugepaged_min_free_kbytes_update(void)
2005 {
2006         mutex_lock(&khugepaged_mutex);
2007         if (khugepaged_enabled() && khugepaged_thread)
2008                 set_recommended_min_free_kbytes();
2009         mutex_unlock(&khugepaged_mutex);
2010 }