mm/khugepaged.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   3
   4 #include <linux/mm.h>
   5 #include <linux/sched.h>
   6 #include <linux/sched/mm.h>
   7 #include <linux/sched/coredump.h>
   8 #include <linux/mmu_notifier.h>
   9 #include <linux/rmap.h>
  10 #include <linux/swap.h>
  11 #include <linux/mm_inline.h>
  12 #include <linux/kthread.h>
  13 #include <linux/khugepaged.h>
  14 #include <linux/freezer.h>
  15 #include <linux/mman.h>
  16 #include <linux/hashtable.h>
  17 #include <linux/userfaultfd_k.h>
  18 #include <linux/page_idle.h>
  19 #include <linux/swapops.h>
  20 #include <linux/shmem_fs.h>
  21
  22 #include <asm/tlb.h>
  23 #include <asm/pgalloc.h>
  24 #include "internal.h"
  25
  26 enum scan_result {
  27         SCAN_FAIL,
  28         SCAN_SUCCEED,
  29         SCAN_PMD_NULL,
  30         SCAN_EXCEED_NONE_PTE,
  31         SCAN_PTE_NON_PRESENT,
  32         SCAN_PAGE_RO,
  33         SCAN_LACK_REFERENCED_PAGE,
  34         SCAN_PAGE_NULL,
  35         SCAN_SCAN_ABORT,
  36         SCAN_PAGE_COUNT,
  37         SCAN_PAGE_LRU,
  38         SCAN_PAGE_LOCK,
  39         SCAN_PAGE_ANON,
  40         SCAN_PAGE_COMPOUND,
  41         SCAN_ANY_PROCESS,
  42         SCAN_VMA_NULL,
  43         SCAN_VMA_CHECK,
  44         SCAN_ADDRESS_RANGE,
  45         SCAN_SWAP_CACHE_PAGE,
  46         SCAN_DEL_PAGE_LRU,
  47         SCAN_ALLOC_HUGE_PAGE_FAIL,
  48         SCAN_CGROUP_CHARGE_FAIL,
  49         SCAN_EXCEED_SWAP_PTE,
  50         SCAN_TRUNCATED,
  51         SCAN_PAGE_HAS_PRIVATE,
  52 };
  53
  54 #define CREATE_TRACE_POINTS
  55 #include <trace/events/huge_memory.h>
  56
  57 /* default scan 8*512 pte (or vmas) every 30 second */
  58 static unsigned int khugepaged_pages_to_scan __read_mostly;
  59 static unsigned int khugepaged_pages_collapsed;
  60 static unsigned int khugepaged_full_scans;
  61 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  62 /* during fragmentation poll the hugepage allocator once every minute */
  63 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  64 static unsigned long khugepaged_sleep_expire;
  65 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  66 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  67 /*
  68  * default collapse hugepages if there is at least one pte mapped like
  69  * it would have happened if the vma was large enough during page
  70  * fault.
  71  */
  72 static unsigned int khugepaged_max_ptes_none __read_mostly;
  73 static unsigned int khugepaged_max_ptes_swap __read_mostly;
  74
  75 #define MM_SLOTS_HASH_BITS 10
  76 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  77
  78 static struct kmem_cache *mm_slot_cache __read_mostly;
  79
  80 /**
  81  * struct mm_slot - hash lookup from mm to mm_slot
  82  * @hash: hash collision list
  83  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  84  * @mm: the mm that this information is valid for
  85  */
  86 struct mm_slot {
  87         struct hlist_node hash;
  88         struct list_head mm_node;
  89         struct mm_struct *mm;
  90 };
  91
  92 /**
  93  * struct khugepaged_scan - cursor for scanning
  94  * @mm_head: the head of the mm list to scan
  95  * @mm_slot: the current mm_slot we are scanning
  96  * @address: the next address inside that to be scanned
  97  *
  98  * There is only the one khugepaged_scan instance of this cursor structure.
  99  */
 100 struct khugepaged_scan {
 101         struct list_head mm_head;
 102         struct mm_slot *mm_slot;
 103         unsigned long address;
 104 };
 105
 106 static struct khugepaged_scan khugepaged_scan = {
 107         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 108 };
 109
 110 #ifdef CONFIG_SYSFS
 111 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 112                                          struct kobj_attribute *attr,
 113                                          char *buf)
 114 {
 115         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 116 }
 117
 118 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 119                                           struct kobj_attribute *attr,
 120                                           const char *buf, size_t count)
 121 {
 122         unsigned long msecs;
 123         int err;
 124
 125         err = kstrtoul(buf, 10, &msecs);
 126         if (err || msecs > UINT_MAX)
 127                 return -EINVAL;
 128
 129         khugepaged_scan_sleep_millisecs = msecs;
 130         khugepaged_sleep_expire = 0;
 131         wake_up_interruptible(&khugepaged_wait);
 132
 133         return count;
 134 }
 135 static struct kobj_attribute scan_sleep_millisecs_attr =
 136         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 137                scan_sleep_millisecs_store);
 138
 139 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 140                                           struct kobj_attribute *attr,
 141                                           char *buf)
 142 {
 143         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 144 }
 145
 146 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 147                                            struct kobj_attribute *attr,
 148                                            const char *buf, size_t count)
 149 {
 150         unsigned long msecs;
 151         int err;
 152
 153         err = kstrtoul(buf, 10, &msecs);
 154         if (err || msecs > UINT_MAX)
 155                 return -EINVAL;
 156
 157         khugepaged_alloc_sleep_millisecs = msecs;
 158         khugepaged_sleep_expire = 0;
 159         wake_up_interruptible(&khugepaged_wait);
 160
 161         return count;
 162 }
 163 static struct kobj_attribute alloc_sleep_millisecs_attr =
 164         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 165                alloc_sleep_millisecs_store);
 166
 167 static ssize_t pages_to_scan_show(struct kobject *kobj,
 168                                   struct kobj_attribute *attr,
 169                                   char *buf)
 170 {
 171         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 172 }
 173 static ssize_t pages_to_scan_store(struct kobject *kobj,
 174                                    struct kobj_attribute *attr,
 175                                    const char *buf, size_t count)
 176 {
 177         int err;
 178         unsigned long pages;
 179
 180         err = kstrtoul(buf, 10, &pages);
 181         if (err || !pages || pages > UINT_MAX)
 182                 return -EINVAL;
 183
 184         khugepaged_pages_to_scan = pages;
 185
 186         return count;
 187 }
 188 static struct kobj_attribute pages_to_scan_attr =
 189         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 190                pages_to_scan_store);
 191
 192 static ssize_t pages_collapsed_show(struct kobject *kobj,
 193                                     struct kobj_attribute *attr,
 194                                     char *buf)
 195 {
 196         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 197 }
 198 static struct kobj_attribute pages_collapsed_attr =
 199         __ATTR_RO(pages_collapsed);
 200
 201 static ssize_t full_scans_show(struct kobject *kobj,
 202                                struct kobj_attribute *attr,
 203                                char *buf)
 204 {
 205         return sprintf(buf, "%u\n", khugepaged_full_scans);
 206 }
 207 static struct kobj_attribute full_scans_attr =
 208         __ATTR_RO(full_scans);
 209
 210 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 211                                       struct kobj_attribute *attr, char *buf)
 212 {
 213         return single_hugepage_flag_show(kobj, attr, buf,
 214                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 215 }
 216 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 217                                        struct kobj_attribute *attr,
 218                                        const char *buf, size_t count)
 219 {
 220         return single_hugepage_flag_store(kobj, attr, buf, count,
 221                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 222 }
 223 static struct kobj_attribute khugepaged_defrag_attr =
 224         __ATTR(defrag, 0644, khugepaged_defrag_show,
 225                khugepaged_defrag_store);
 226
 227 /*
 228  * max_ptes_none controls if khugepaged should collapse hugepages over
 229  * any unmapped ptes in turn potentially increasing the memory
 230  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 231  * reduce the available free memory in the system as it
 232  * runs. Increasing max_ptes_none will instead potentially reduce the
 233  * free memory in the system during the khugepaged scan.
 234  */
 235 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 236                                              struct kobj_attribute *attr,
 237                                              char *buf)
 238 {
 239         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 240 }
 241 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 242                                               struct kobj_attribute *attr,
 243                                               const char *buf, size_t count)
 244 {
 245         int err;
 246         unsigned long max_ptes_none;
 247
 248         err = kstrtoul(buf, 10, &max_ptes_none);
 249         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 250                 return -EINVAL;
 251
 252         khugepaged_max_ptes_none = max_ptes_none;
 253
 254         return count;
 255 }
 256 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 257         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 258                khugepaged_max_ptes_none_store);
 259
 260 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
 261                                              struct kobj_attribute *attr,
 262                                              char *buf)
 263 {
 264         return sprintf(buf, "%u\n", khugepaged_max_ptes_swap);
 265 }
 266
 267 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
 268                                               struct kobj_attribute *attr,
 269                                               const char *buf, size_t count)
 270 {
 271         int err;
 272         unsigned long max_ptes_swap;
 273
 274         err  = kstrtoul(buf, 10, &max_ptes_swap);
 275         if (err || max_ptes_swap > HPAGE_PMD_NR-1)
 276                 return -EINVAL;
 277
 278         khugepaged_max_ptes_swap = max_ptes_swap;
 279
 280         return count;
 281 }
 282
 283 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
 284         __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
 285                khugepaged_max_ptes_swap_store);
 286
 287 static struct attribute *khugepaged_attr[] = {
 288         &khugepaged_defrag_attr.attr,
 289         &khugepaged_max_ptes_none_attr.attr,
 290         &pages_to_scan_attr.attr,
 291         &pages_collapsed_attr.attr,
 292         &full_scans_attr.attr,
 293         &scan_sleep_millisecs_attr.attr,
 294         &alloc_sleep_millisecs_attr.attr,
 295         &khugepaged_max_ptes_swap_attr.attr,
 296         NULL,
 297 };
 298
 299 struct attribute_group khugepaged_attr_group = {
 300         .attrs = khugepaged_attr,
 301         .name = "khugepaged",
 302 };
 303 #endif /* CONFIG_SYSFS */
 304
 305 #define VM_NO_KHUGEPAGED (VM_SPECIAL | VM_HUGETLB)
 306
 307 int hugepage_madvise(struct vm_area_struct *vma,
 308                      unsigned long *vm_flags, int advice)
 309 {
 310         switch (advice) {
 311         case MADV_HUGEPAGE:
 312 #ifdef CONFIG_S390
 313                 /*
 314                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
 315                  * can't handle this properly after s390_enable_sie, so we simply
 316                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
 317                  */
 318                 if (mm_has_pgste(vma->vm_mm))
 319                         return 0;
 320 #endif
 321                 *vm_flags &= ~VM_NOHUGEPAGE;
 322                 *vm_flags |= VM_HUGEPAGE;
 323                 /*
 324                  * If the vma become good for khugepaged to scan,
 325                  * register it here without waiting a page fault that
 326                  * may not happen any time soon.
 327                  */
 328                 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
 329                                 khugepaged_enter_vma_merge(vma, *vm_flags))
 330                         return -ENOMEM;
 331                 break;
 332         case MADV_NOHUGEPAGE:
 333                 *vm_flags &= ~VM_HUGEPAGE;
 334                 *vm_flags |= VM_NOHUGEPAGE;
 335                 /*
 336                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
 337                  * this vma even if we leave the mm registered in khugepaged if
 338                  * it got registered before VM_NOHUGEPAGE was set.
 339                  */
 340                 break;
 341         }
 342
 343         return 0;
 344 }
 345
 346 int __init khugepaged_init(void)
 347 {
 348         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
 349                                           sizeof(struct mm_slot),
 350                                           __alignof__(struct mm_slot), 0, NULL);
 351         if (!mm_slot_cache)
 352                 return -ENOMEM;
 353
 354         khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
 355         khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
 356         khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
 357
 358         return 0;
 359 }
 360
 361 void __init khugepaged_destroy(void)
 362 {
 363         kmem_cache_destroy(mm_slot_cache);
 364 }
 365
 366 static inline struct mm_slot *alloc_mm_slot(void)
 367 {
 368         if (!mm_slot_cache)     /* initialization failed */
 369                 return NULL;
 370         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
 371 }
 372
 373 static inline void free_mm_slot(struct mm_slot *mm_slot)
 374 {
 375         kmem_cache_free(mm_slot_cache, mm_slot);
 376 }
 377
 378 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
 379 {
 380         struct mm_slot *mm_slot;
 381
 382         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
 383                 if (mm == mm_slot->mm)
 384                         return mm_slot;
 385
 386         return NULL;
 387 }
 388
 389 static void insert_to_mm_slots_hash(struct mm_struct *mm,
 390                                     struct mm_slot *mm_slot)
 391 {
 392         mm_slot->mm = mm;
 393         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
 394 }
 395
 396 static inline int khugepaged_test_exit(struct mm_struct *mm)
 397 {
 398         return atomic_read(&mm->mm_users) == 0;
 399 }
 400
 401 static bool hugepage_vma_check(struct vm_area_struct *vma,
 402                                unsigned long vm_flags)
 403 {
 404         if ((!(vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
 405             (vm_flags & VM_NOHUGEPAGE) ||
 406             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
 407                 return false;
 408
 409         if (shmem_file(vma->vm_file) ||
 410             (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) &&
 411              vma->vm_file &&
 412              (vm_flags & VM_DENYWRITE))) {
 413                 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
 414                         return false;
 415                 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
 416                                 HPAGE_PMD_NR);
 417         }
 418         if (!vma->anon_vma || vma->vm_ops)
 419                 return false;
 420         if (is_vma_temporary_stack(vma))
 421                 return false;
 422         return !(vm_flags & VM_NO_KHUGEPAGED);
 423 }
 424
 425 int __khugepaged_enter(struct mm_struct *mm)
 426 {
 427         struct mm_slot *mm_slot;
 428         int wakeup;
 429
 430         mm_slot = alloc_mm_slot();
 431         if (!mm_slot)
 432                 return -ENOMEM;
 433
 434         /* __khugepaged_exit() must not run from under us */
 435         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
 436         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
 437                 free_mm_slot(mm_slot);
 438                 return 0;
 439         }
 440
 441         spin_lock(&khugepaged_mm_lock);
 442         insert_to_mm_slots_hash(mm, mm_slot);
 443         /*
 444          * Insert just behind the scanning cursor, to let the area settle
 445          * down a little.
 446          */
 447         wakeup = list_empty(&khugepaged_scan.mm_head);
 448         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
 449         spin_unlock(&khugepaged_mm_lock);
 450
 451         mmgrab(mm);
 452         if (wakeup)
 453                 wake_up_interruptible(&khugepaged_wait);
 454
 455         return 0;
 456 }
 457
 458 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
 459                                unsigned long vm_flags)
 460 {
 461         unsigned long hstart, hend;
 462
 463         /*
 464          * khugepaged only supports read-only files for non-shmem files.
 465          * khugepaged does not yet work on special mappings. And
 466          * file-private shmem THP is not supported.
 467          */
 468         if (!hugepage_vma_check(vma, vm_flags))
 469                 return 0;
 470
 471         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 472         hend = vma->vm_end & HPAGE_PMD_MASK;
 473         if (hstart < hend)
 474                 return khugepaged_enter(vma, vm_flags);
 475         return 0;
 476 }
 477
 478 void __khugepaged_exit(struct mm_struct *mm)
 479 {
 480         struct mm_slot *mm_slot;
 481         int free = 0;
 482
 483         spin_lock(&khugepaged_mm_lock);
 484         mm_slot = get_mm_slot(mm);
 485         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
 486                 hash_del(&mm_slot->hash);
 487                 list_del(&mm_slot->mm_node);
 488                 free = 1;
 489         }
 490         spin_unlock(&khugepaged_mm_lock);
 491
 492         if (free) {
 493                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
 494                 free_mm_slot(mm_slot);
 495                 mmdrop(mm);
 496         } else if (mm_slot) {
 497                 /*
 498                  * This is required to serialize against
 499                  * khugepaged_test_exit() (which is guaranteed to run
 500                  * under mmap sem read mode). Stop here (after we
 501                  * return all pagetables will be destroyed) until
 502                  * khugepaged has finished working on the pagetables
 503                  * under the mmap_sem.
 504                  */
 505                 down_write(&mm->mmap_sem);
 506                 up_write(&mm->mmap_sem);
 507         }
 508 }
 509
 510 static void release_pte_page(struct page *page)
 511 {
 512         dec_node_page_state(page, NR_ISOLATED_ANON + page_is_file_cache(page));
 513         unlock_page(page);
 514         putback_lru_page(page);
 515 }
 516
 517 static void release_pte_pages(pte_t *pte, pte_t *_pte)
 518 {
 519         while (--_pte >= pte) {
 520                 pte_t pteval = *_pte;
 521                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
 522                         release_pte_page(pte_page(pteval));
 523         }
 524 }
 525
 526 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
 527                                         unsigned long address,
 528                                         pte_t *pte)
 529 {
 530         struct page *page = NULL;
 531         pte_t *_pte;
 532         int none_or_zero = 0, result = 0, referenced = 0;
 533         bool writable = false;
 534
 535         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
 536              _pte++, address += PAGE_SIZE) {
 537                 pte_t pteval = *_pte;
 538                 if (pte_none(pteval) || (pte_present(pteval) &&
 539                                 is_zero_pfn(pte_pfn(pteval)))) {
 540                         if (!userfaultfd_armed(vma) &&
 541                             ++none_or_zero <= khugepaged_max_ptes_none) {
 542                                 continue;
 543                         } else {
 544                                 result = SCAN_EXCEED_NONE_PTE;
 545                                 goto out;
 546                         }
 547                 }
 548                 if (!pte_present(pteval)) {
 549                         result = SCAN_PTE_NON_PRESENT;
 550                         goto out;
 551                 }
 552                 page = vm_normal_page(vma, address, pteval);
 553                 if (unlikely(!page)) {
 554                         result = SCAN_PAGE_NULL;
 555                         goto out;
 556                 }
 557
 558                 /* TODO: teach khugepaged to collapse THP mapped with pte */
 559                 if (PageCompound(page)) {
 560                         result = SCAN_PAGE_COMPOUND;
 561                         goto out;
 562                 }
 563
 564                 VM_BUG_ON_PAGE(!PageAnon(page), page);
 565
 566                 /*
 567                  * We can do it before isolate_lru_page because the
 568                  * page can't be freed from under us. NOTE: PG_lock
 569                  * is needed to serialize against split_huge_page
 570                  * when invoked from the VM.
 571                  */
 572                 if (!trylock_page(page)) {
 573                         result = SCAN_PAGE_LOCK;
 574                         goto out;
 575                 }
 576
 577                 /*
 578                  * cannot use mapcount: can't collapse if there's a gup pin.
 579                  * The page must only be referenced by the scanned process
 580                  * and page swap cache.
 581                  */
 582                 if (page_count(page) != 1 + PageSwapCache(page)) {
 583                         unlock_page(page);
 584                         result = SCAN_PAGE_COUNT;
 585                         goto out;
 586                 }
 587                 if (pte_write(pteval)) {
 588                         writable = true;
 589                 } else {
 590                         if (PageSwapCache(page) &&
 591                             !reuse_swap_page(page, NULL)) {
 592                                 unlock_page(page);
 593                                 result = SCAN_SWAP_CACHE_PAGE;
 594                                 goto out;
 595                         }
 596                         /*
 597                          * Page is not in the swap cache. It can be collapsed
 598                          * into a THP.
 599                          */
 600                 }
 601
 602                 /*
 603                  * Isolate the page to avoid collapsing an hugepage
 604                  * currently in use by the VM.
 605                  */
 606                 if (isolate_lru_page(page)) {
 607                         unlock_page(page);
 608                         result = SCAN_DEL_PAGE_LRU;
 609                         goto out;
 610                 }
 611                 inc_node_page_state(page,
 612                                 NR_ISOLATED_ANON + page_is_file_cache(page));
 613                 VM_BUG_ON_PAGE(!PageLocked(page), page);
 614                 VM_BUG_ON_PAGE(PageLRU(page), page);
 615
 616                 /* There should be enough young pte to collapse the page */
 617                 if (pte_young(pteval) ||
 618                     page_is_young(page) || PageReferenced(page) ||
 619                     mmu_notifier_test_young(vma->vm_mm, address))
 620                         referenced++;
 621         }
 622         if (likely(writable)) {
 623                 if (likely(referenced)) {
 624                         result = SCAN_SUCCEED;
 625                         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 626                                                             referenced, writable, result);
 627                         return 1;
 628                 }
 629         } else {
 630                 result = SCAN_PAGE_RO;
 631         }
 632
 633 out:
 634         release_pte_pages(pte, _pte);
 635         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 636                                             referenced, writable, result);
 637         return 0;
 638 }
 639
 640 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
 641                                       struct vm_area_struct *vma,
 642                                       unsigned long address,
 643                                       spinlock_t *ptl)
 644 {
 645         pte_t *_pte;
 646         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 647                                 _pte++, page++, address += PAGE_SIZE) {
 648                 pte_t pteval = *_pte;
 649                 struct page *src_page;
 650
 651                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
 652                         clear_user_highpage(page, address);
 653                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
 654                         if (is_zero_pfn(pte_pfn(pteval))) {
 655                                 /*
 656                                  * ptl mostly unnecessary.
 657                                  */
 658                                 spin_lock(ptl);
 659                                 /*
 660                                  * paravirt calls inside pte_clear here are
 661                                  * superfluous.
 662                                  */
 663                                 pte_clear(vma->vm_mm, address, _pte);
 664                                 spin_unlock(ptl);
 665                         }
 666                 } else {
 667                         src_page = pte_page(pteval);
 668                         copy_user_highpage(page, src_page, address, vma);
 669                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
 670                         release_pte_page(src_page);
 671                         /*
 672                          * ptl mostly unnecessary, but preempt has to
 673                          * be disabled to update the per-cpu stats
 674                          * inside page_remove_rmap().
 675                          */
 676                         spin_lock(ptl);
 677                         /*
 678                          * paravirt calls inside pte_clear here are
 679                          * superfluous.
 680                          */
 681                         pte_clear(vma->vm_mm, address, _pte);
 682                         page_remove_rmap(src_page, false);
 683                         spin_unlock(ptl);
 684                         free_page_and_swap_cache(src_page);
 685                 }
 686         }
 687 }
 688
 689 static void khugepaged_alloc_sleep(void)
 690 {
 691         DEFINE_WAIT(wait);
 692
 693         add_wait_queue(&khugepaged_wait, &wait);
 694         freezable_schedule_timeout_interruptible(
 695                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
 696         remove_wait_queue(&khugepaged_wait, &wait);
 697 }
 698
 699 static int khugepaged_node_load[MAX_NUMNODES];
 700
 701 static bool khugepaged_scan_abort(int nid)
 702 {
 703         int i;
 704
 705         /*
 706          * If node_reclaim_mode is disabled, then no extra effort is made to
 707          * allocate memory locally.
 708          */
 709         if (!node_reclaim_mode)
 710                 return false;
 711
 712         /* If there is a count for this node already, it must be acceptable */
 713         if (khugepaged_node_load[nid])
 714                 return false;
 715
 716         for (i = 0; i < MAX_NUMNODES; i++) {
 717                 if (!khugepaged_node_load[i])
 718                         continue;
 719                 if (node_distance(nid, i) > node_reclaim_distance)
 720                         return true;
 721         }
 722         return false;
 723 }
 724
 725 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
 726 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
 727 {
 728         return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
 729 }
 730
 731 #ifdef CONFIG_NUMA
 732 static int khugepaged_find_target_node(void)
 733 {
 734         static int last_khugepaged_target_node = NUMA_NO_NODE;
 735         int nid, target_node = 0, max_value = 0;
 736
 737         /* find first node with max normal pages hit */
 738         for (nid = 0; nid < MAX_NUMNODES; nid++)
 739                 if (khugepaged_node_load[nid] > max_value) {
 740                         max_value = khugepaged_node_load[nid];
 741                         target_node = nid;
 742                 }
 743
 744         /* do some balance if several nodes have the same hit record */
 745         if (target_node <= last_khugepaged_target_node)
 746                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
 747                                 nid++)
 748                         if (max_value == khugepaged_node_load[nid]) {
 749                                 target_node = nid;
 750                                 break;
 751                         }
 752
 753         last_khugepaged_target_node = target_node;
 754         return target_node;
 755 }
 756
 757 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 758 {
 759         if (IS_ERR(*hpage)) {
 760                 if (!*wait)
 761                         return false;
 762
 763                 *wait = false;
 764                 *hpage = NULL;
 765                 khugepaged_alloc_sleep();
 766         } else if (*hpage) {
 767                 put_page(*hpage);
 768                 *hpage = NULL;
 769         }
 770
 771         return true;
 772 }
 773
 774 static struct page *
 775 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 776 {
 777         VM_BUG_ON_PAGE(*hpage, *hpage);
 778
 779         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
 780         if (unlikely(!*hpage)) {
 781                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 782                 *hpage = ERR_PTR(-ENOMEM);
 783                 return NULL;
 784         }
 785
 786         prep_transhuge_page(*hpage);
 787         count_vm_event(THP_COLLAPSE_ALLOC);
 788         return *hpage;
 789 }
 790 #else
 791 static int khugepaged_find_target_node(void)
 792 {
 793         return 0;
 794 }
 795
 796 static inline struct page *alloc_khugepaged_hugepage(void)
 797 {
 798         struct page *page;
 799
 800         page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
 801                            HPAGE_PMD_ORDER);
 802         if (page)
 803                 prep_transhuge_page(page);
 804         return page;
 805 }
 806
 807 static struct page *khugepaged_alloc_hugepage(bool *wait)
 808 {
 809         struct page *hpage;
 810
 811         do {
 812                 hpage = alloc_khugepaged_hugepage();
 813                 if (!hpage) {
 814                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 815                         if (!*wait)
 816                                 return NULL;
 817
 818                         *wait = false;
 819                         khugepaged_alloc_sleep();
 820                 } else
 821                         count_vm_event(THP_COLLAPSE_ALLOC);
 822         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
 823
 824         return hpage;
 825 }
 826
 827 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 828 {
 829         if (!*hpage)
 830                 *hpage = khugepaged_alloc_hugepage(wait);
 831
 832         if (unlikely(!*hpage))
 833                 return false;
 834
 835         return true;
 836 }
 837
 838 static struct page *
 839 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 840 {
 841         VM_BUG_ON(!*hpage);
 842
 843         return  *hpage;
 844 }
 845 #endif
 846
 847 /*
 848  * If mmap_sem temporarily dropped, revalidate vma
 849  * before taking mmap_sem.
 850  * Return 0 if succeeds, otherwise return none-zero
 851  * value (scan code).
 852  */
 853
 854 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
 855                 struct vm_area_struct **vmap)
 856 {
 857         struct vm_area_struct *vma;
 858         unsigned long hstart, hend;
 859
 860         if (unlikely(khugepaged_test_exit(mm)))
 861                 return SCAN_ANY_PROCESS;
 862
 863         *vmap = vma = find_vma(mm, address);
 864         if (!vma)
 865                 return SCAN_VMA_NULL;
 866
 867         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 868         hend = vma->vm_end & HPAGE_PMD_MASK;
 869         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
 870                 return SCAN_ADDRESS_RANGE;
 871         if (!hugepage_vma_check(vma, vma->vm_flags))
 872                 return SCAN_VMA_CHECK;
 873         return 0;
 874 }
 875
 876 /*
 877  * Bring missing pages in from swap, to complete THP collapse.
 878  * Only done if khugepaged_scan_pmd believes it is worthwhile.
 879  *
 880  * Called and returns without pte mapped or spinlocks held,
 881  * but with mmap_sem held to protect against vma changes.
 882  */
 883
 884 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
 885                                         struct vm_area_struct *vma,
 886                                         unsigned long address, pmd_t *pmd,
 887                                         int referenced)
 888 {
 889         int swapped_in = 0;
 890         vm_fault_t ret = 0;
 891         struct vm_fault vmf = {
 892                 .vma = vma,
 893                 .address = address,
 894                 .flags = FAULT_FLAG_ALLOW_RETRY,
 895                 .pmd = pmd,
 896                 .pgoff = linear_page_index(vma, address),
 897         };
 898
 899         /* we only decide to swapin, if there is enough young ptes */
 900         if (referenced < HPAGE_PMD_NR/2) {
 901                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 902                 return false;
 903         }
 904         vmf.pte = pte_offset_map(pmd, address);
 905         for (; vmf.address < address + HPAGE_PMD_NR*PAGE_SIZE;
 906                         vmf.pte++, vmf.address += PAGE_SIZE) {
 907                 vmf.orig_pte = *vmf.pte;
 908                 if (!is_swap_pte(vmf.orig_pte))
 909                         continue;
 910                 swapped_in++;
 911                 ret = do_swap_page(&vmf);
 912
 913                 /* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */
 914                 if (ret & VM_FAULT_RETRY) {
 915                         down_read(&mm->mmap_sem);
 916                         if (hugepage_vma_revalidate(mm, address, &vmf.vma)) {
 917                                 /* vma is no longer available, don't continue to swapin */
 918                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 919                                 return false;
 920                         }
 921                         /* check if the pmd is still valid */
 922                         if (mm_find_pmd(mm, address) != pmd) {
 923                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 924                                 return false;
 925                         }
 926                 }
 927                 if (ret & VM_FAULT_ERROR) {
 928                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 929                         return false;
 930                 }
 931                 /* pte is unmapped now, we need to map it */
 932                 vmf.pte = pte_offset_map(pmd, vmf.address);
 933         }
 934         vmf.pte--;
 935         pte_unmap(vmf.pte);
 936         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
 937         return true;
 938 }
 939
 940 static void collapse_huge_page(struct mm_struct *mm,
 941                                    unsigned long address,
 942                                    struct page **hpage,
 943                                    int node, int referenced)
 944 {
 945         pmd_t *pmd, _pmd;
 946         pte_t *pte;
 947         pgtable_t pgtable;
 948         struct page *new_page;
 949         spinlock_t *pmd_ptl, *pte_ptl;
 950         int isolated = 0, result = 0;
 951         struct mem_cgroup *memcg;
 952         struct vm_area_struct *vma;
 953         struct mmu_notifier_range range;
 954         gfp_t gfp;
 955
 956         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 957
 958         /* Only allocate from the target node */
 959         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
 960
 961         /*
 962          * Before allocating the hugepage, release the mmap_sem read lock.
 963          * The allocation can take potentially a long time if it involves
 964          * sync compaction, and we do not need to hold the mmap_sem during
 965          * that. We will recheck the vma after taking it again in write mode.
 966          */
 967         up_read(&mm->mmap_sem);
 968         new_page = khugepaged_alloc_page(hpage, gfp, node);
 969         if (!new_page) {
 970                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
 971                 goto out_nolock;
 972         }
 973
 974         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
 975                 result = SCAN_CGROUP_CHARGE_FAIL;
 976                 goto out_nolock;
 977         }
 978
 979         down_read(&mm->mmap_sem);
 980         result = hugepage_vma_revalidate(mm, address, &vma);
 981         if (result) {
 982                 mem_cgroup_cancel_charge(new_page, memcg, true);
 983                 up_read(&mm->mmap_sem);
 984                 goto out_nolock;
 985         }
 986
 987         pmd = mm_find_pmd(mm, address);
 988         if (!pmd) {
 989                 result = SCAN_PMD_NULL;
 990                 mem_cgroup_cancel_charge(new_page, memcg, true);
 991                 up_read(&mm->mmap_sem);
 992                 goto out_nolock;
 993         }
 994
 995         /*
 996          * __collapse_huge_page_swapin always returns with mmap_sem locked.
 997          * If it fails, we release mmap_sem and jump out_nolock.
 998          * Continuing to collapse causes inconsistency.
 999          */
1000         if (!__collapse_huge_page_swapin(mm, vma, address, pmd, referenced)) {
1001                 mem_cgroup_cancel_charge(new_page, memcg, true);
1002                 up_read(&mm->mmap_sem);
1003                 goto out_nolock;
1004         }
1005
1006         up_read(&mm->mmap_sem);
1007         /*
1008          * Prevent all access to pagetables with the exception of
1009          * gup_fast later handled by the ptep_clear_flush and the VM
1010          * handled by the anon_vma lock + PG_lock.
1011          */
1012         down_write(&mm->mmap_sem);
1013         result = SCAN_ANY_PROCESS;
1014         if (!mmget_still_valid(mm))
1015                 goto out;
1016         result = hugepage_vma_revalidate(mm, address, &vma);
1017         if (result)
1018                 goto out;
1019         /* check if the pmd is still valid */
1020         if (mm_find_pmd(mm, address) != pmd)
1021                 goto out;
1022
1023         anon_vma_lock_write(vma->anon_vma);
1024
1025         pte = pte_offset_map(pmd, address);
1026         pte_ptl = pte_lockptr(mm, pmd);
1027
1028         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1029                                 address, address + HPAGE_PMD_SIZE);
1030         mmu_notifier_invalidate_range_start(&range);
1031         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1032         /*
1033          * After this gup_fast can't run anymore. This also removes
1034          * any huge TLB entry from the CPU so we won't allow
1035          * huge and small TLB entries for the same virtual address
1036          * to avoid the risk of CPU bugs in that area.
1037          */
1038         _pmd = pmdp_collapse_flush(vma, address, pmd);
1039         spin_unlock(pmd_ptl);
1040         mmu_notifier_invalidate_range_end(&range);
1041
1042         spin_lock(pte_ptl);
1043         isolated = __collapse_huge_page_isolate(vma, address, pte);
1044         spin_unlock(pte_ptl);
1045
1046         if (unlikely(!isolated)) {
1047                 pte_unmap(pte);
1048                 spin_lock(pmd_ptl);
1049                 BUG_ON(!pmd_none(*pmd));
1050                 /*
1051                  * We can only use set_pmd_at when establishing
1052                  * hugepmds and never for establishing regular pmds that
1053                  * points to regular pagetables. Use pmd_populate for that
1054                  */
1055                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1056                 spin_unlock(pmd_ptl);
1057                 anon_vma_unlock_write(vma->anon_vma);
1058                 result = SCAN_FAIL;
1059                 goto out;
1060         }
1061
1062         /*
1063          * All pages are isolated and locked so anon_vma rmap
1064          * can't run anymore.
1065          */
1066         anon_vma_unlock_write(vma->anon_vma);
1067
1068         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
1069         pte_unmap(pte);
1070         __SetPageUptodate(new_page);
1071         pgtable = pmd_pgtable(_pmd);
1072
1073         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1074         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1075
1076         /*
1077          * spin_lock() below is not the equivalent of smp_wmb(), so
1078          * this is needed to avoid the copy_huge_page writes to become
1079          * visible after the set_pmd_at() write.
1080          */
1081         smp_wmb();
1082
1083         spin_lock(pmd_ptl);
1084         BUG_ON(!pmd_none(*pmd));
1085         page_add_new_anon_rmap(new_page, vma, address, true);
1086         mem_cgroup_commit_charge(new_page, memcg, false, true);
1087         count_memcg_events(memcg, THP_COLLAPSE_ALLOC, 1);
1088         lru_cache_add_active_or_unevictable(new_page, vma);
1089         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1090         set_pmd_at(mm, address, pmd, _pmd);
1091         update_mmu_cache_pmd(vma, address, pmd);
1092         spin_unlock(pmd_ptl);
1093
1094         *hpage = NULL;
1095
1096         khugepaged_pages_collapsed++;
1097         result = SCAN_SUCCEED;
1098 out_up_write:
1099         up_write(&mm->mmap_sem);
1100 out_nolock:
1101         trace_mm_collapse_huge_page(mm, isolated, result);
1102         return;
1103 out:
1104         mem_cgroup_cancel_charge(new_page, memcg, true);
1105         goto out_up_write;
1106 }
1107
1108 static int khugepaged_scan_pmd(struct mm_struct *mm,
1109                                struct vm_area_struct *vma,
1110                                unsigned long address,
1111                                struct page **hpage)
1112 {
1113         pmd_t *pmd;
1114         pte_t *pte, *_pte;
1115         int ret = 0, none_or_zero = 0, result = 0, referenced = 0;
1116         struct page *page = NULL;
1117         unsigned long _address;
1118         spinlock_t *ptl;
1119         int node = NUMA_NO_NODE, unmapped = 0;
1120         bool writable = false;
1121
1122         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1123
1124         pmd = mm_find_pmd(mm, address);
1125         if (!pmd) {
1126                 result = SCAN_PMD_NULL;
1127                 goto out;
1128         }
1129
1130         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1131         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1132         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1133              _pte++, _address += PAGE_SIZE) {
1134                 pte_t pteval = *_pte;
1135                 if (is_swap_pte(pteval)) {
1136                         if (++unmapped <= khugepaged_max_ptes_swap) {
1137                                 continue;
1138                         } else {
1139                                 result = SCAN_EXCEED_SWAP_PTE;
1140                                 goto out_unmap;
1141                         }
1142                 }
1143                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1144                         if (!userfaultfd_armed(vma) &&
1145                             ++none_or_zero <= khugepaged_max_ptes_none) {
1146                                 continue;
1147                         } else {
1148                                 result = SCAN_EXCEED_NONE_PTE;
1149                                 goto out_unmap;
1150                         }
1151                 }
1152                 if (!pte_present(pteval)) {
1153                         result = SCAN_PTE_NON_PRESENT;
1154                         goto out_unmap;
1155                 }
1156                 if (pte_write(pteval))
1157                         writable = true;
1158
1159                 page = vm_normal_page(vma, _address, pteval);
1160                 if (unlikely(!page)) {
1161                         result = SCAN_PAGE_NULL;
1162                         goto out_unmap;
1163                 }
1164
1165                 /* TODO: teach khugepaged to collapse THP mapped with pte */
1166                 if (PageCompound(page)) {
1167                         result = SCAN_PAGE_COMPOUND;
1168                         goto out_unmap;
1169                 }
1170
1171                 /*
1172                  * Record which node the original page is from and save this
1173                  * information to khugepaged_node_load[].
1174                  * Khupaged will allocate hugepage from the node has the max
1175                  * hit record.
1176                  */
1177                 node = page_to_nid(page);
1178                 if (khugepaged_scan_abort(node)) {
1179                         result = SCAN_SCAN_ABORT;
1180                         goto out_unmap;
1181                 }
1182                 khugepaged_node_load[node]++;
1183                 if (!PageLRU(page)) {
1184                         result = SCAN_PAGE_LRU;
1185                         goto out_unmap;
1186                 }
1187                 if (PageLocked(page)) {
1188                         result = SCAN_PAGE_LOCK;
1189                         goto out_unmap;
1190                 }
1191                 if (!PageAnon(page)) {
1192                         result = SCAN_PAGE_ANON;
1193                         goto out_unmap;
1194                 }
1195
1196                 /*
1197                  * cannot use mapcount: can't collapse if there's a gup pin.
1198                  * The page must only be referenced by the scanned process
1199                  * and page swap cache.
1200                  */
1201                 if (page_count(page) != 1 + PageSwapCache(page)) {
1202                         result = SCAN_PAGE_COUNT;
1203                         goto out_unmap;
1204                 }
1205                 if (pte_young(pteval) ||
1206                     page_is_young(page) || PageReferenced(page) ||
1207                     mmu_notifier_test_young(vma->vm_mm, address))
1208                         referenced++;
1209         }
1210         if (writable) {
1211                 if (referenced) {
1212                         result = SCAN_SUCCEED;
1213                         ret = 1;
1214                 } else {
1215                         result = SCAN_LACK_REFERENCED_PAGE;
1216                 }
1217         } else {
1218                 result = SCAN_PAGE_RO;
1219         }
1220 out_unmap:
1221         pte_unmap_unlock(pte, ptl);
1222         if (ret) {
1223                 node = khugepaged_find_target_node();
1224                 /* collapse_huge_page will return with the mmap_sem released */
1225                 collapse_huge_page(mm, address, hpage, node, referenced);
1226         }
1227 out:
1228         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1229                                      none_or_zero, result, unmapped);
1230         return ret;
1231 }
1232
1233 static void collect_mm_slot(struct mm_slot *mm_slot)
1234 {
1235         struct mm_struct *mm = mm_slot->mm;
1236
1237         lockdep_assert_held(&khugepaged_mm_lock);
1238
1239         if (khugepaged_test_exit(mm)) {
1240                 /* free mm_slot */
1241                 hash_del(&mm_slot->hash);
1242                 list_del(&mm_slot->mm_node);
1243
1244                 /*
1245                  * Not strictly needed because the mm exited already.
1246                  *
1247                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1248                  */
1249
1250                 /* khugepaged_mm_lock actually not necessary for the below */
1251                 free_mm_slot(mm_slot);
1252                 mmdrop(mm);
1253         }
1254 }
1255
1256 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
1257 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1258 {
1259         struct vm_area_struct *vma;
1260         unsigned long addr;
1261         pmd_t *pmd, _pmd;
1262
1263         i_mmap_lock_write(mapping);
1264         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1265                 /* probably overkill */
1266                 if (vma->anon_vma)
1267                         continue;
1268                 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1269                 if (addr & ~HPAGE_PMD_MASK)
1270                         continue;
1271                 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1272                         continue;
1273                 pmd = mm_find_pmd(vma->vm_mm, addr);
1274                 if (!pmd)
1275                         continue;
1276                 /*
1277                  * We need exclusive mmap_sem to retract page table.
1278                  * If trylock fails we would end up with pte-mapped THP after
1279                  * re-fault. Not ideal, but it's more important to not disturb
1280                  * the system too much.
1281                  */
1282                 if (down_write_trylock(&vma->vm_mm->mmap_sem)) {
1283                         spinlock_t *ptl = pmd_lock(vma->vm_mm, pmd);
1284                         /* assume page table is clear */
1285                         _pmd = pmdp_collapse_flush(vma, addr, pmd);
1286                         spin_unlock(ptl);
1287                         up_write(&vma->vm_mm->mmap_sem);
1288                         mm_dec_nr_ptes(vma->vm_mm);
1289                         pte_free(vma->vm_mm, pmd_pgtable(_pmd));
1290                 }
1291         }
1292         i_mmap_unlock_write(mapping);
1293 }
1294
1295 /**
1296  * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1297  *
1298  * Basic scheme is simple, details are more complex:
1299  *  - allocate and lock a new huge page;
1300  *  - scan page cache replacing old pages with the new one
1301  *    + swap/gup in pages if necessary;
1302  *    + fill in gaps;
1303  *    + keep old pages around in case rollback is required;
1304  *  - if replacing succeeds:
1305  *    + copy data over;
1306  *    + free old pages;
1307  *    + unlock huge page;
1308  *  - if replacing failed;
1309  *    + put all pages back and unfreeze them;
1310  *    + restore gaps in the page cache;
1311  *    + unlock and free huge page;
1312  */
1313 static void collapse_file(struct mm_struct *mm,
1314                 struct file *file, pgoff_t start,
1315                 struct page **hpage, int node)
1316 {
1317         struct address_space *mapping = file->f_mapping;
1318         gfp_t gfp;
1319         struct page *new_page;
1320         struct mem_cgroup *memcg;
1321         pgoff_t index, end = start + HPAGE_PMD_NR;
1322         LIST_HEAD(pagelist);
1323         XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1324         int nr_none = 0, result = SCAN_SUCCEED;
1325         bool is_shmem = shmem_file(file);
1326
1327         VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1328         VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1329
1330         /* Only allocate from the target node */
1331         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1332
1333         new_page = khugepaged_alloc_page(hpage, gfp, node);
1334         if (!new_page) {
1335                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1336                 goto out;
1337         }
1338
1339         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
1340                 result = SCAN_CGROUP_CHARGE_FAIL;
1341                 goto out;
1342         }
1343
1344         /* This will be less messy when we use multi-index entries */
1345         do {
1346                 xas_lock_irq(&xas);
1347                 xas_create_range(&xas);
1348                 if (!xas_error(&xas))
1349                         break;
1350                 xas_unlock_irq(&xas);
1351                 if (!xas_nomem(&xas, GFP_KERNEL)) {
1352                         mem_cgroup_cancel_charge(new_page, memcg, true);
1353                         result = SCAN_FAIL;
1354                         goto out;
1355                 }
1356         } while (1);
1357
1358         __SetPageLocked(new_page);
1359         if (is_shmem)
1360                 __SetPageSwapBacked(new_page);
1361         new_page->index = start;
1362         new_page->mapping = mapping;
1363
1364         /*
1365          * At this point the new_page is locked and not up-to-date.
1366          * It's safe to insert it into the page cache, because nobody would
1367          * be able to map it or use it in another way until we unlock it.
1368          */
1369
1370         xas_set(&xas, start);
1371         for (index = start; index < end; index++) {
1372                 struct page *page = xas_next(&xas);
1373
1374                 VM_BUG_ON(index != xas.xa_index);
1375                 if (is_shmem) {
1376                         if (!page) {
1377                                 /*
1378                                  * Stop if extent has been truncated or
1379                                  * hole-punched, and is now completely
1380                                  * empty.
1381                                  */
1382                                 if (index == start) {
1383                                         if (!xas_next_entry(&xas, end - 1)) {
1384                                                 result = SCAN_TRUNCATED;
1385                                                 goto xa_locked;
1386                                         }
1387                                         xas_set(&xas, index);
1388                                 }
1389                                 if (!shmem_charge(mapping->host, 1)) {
1390                                         result = SCAN_FAIL;
1391                                         goto xa_locked;
1392                                 }
1393                                 xas_store(&xas, new_page);
1394                                 nr_none++;
1395                                 continue;
1396                         }
1397
1398                         if (xa_is_value(page) || !PageUptodate(page)) {
1399                                 xas_unlock_irq(&xas);
1400                                 /* swap in or instantiate fallocated page */
1401                                 if (shmem_getpage(mapping->host, index, &page,
1402                                                   SGP_NOHUGE)) {
1403                                         result = SCAN_FAIL;
1404                                         goto xa_unlocked;
1405                                 }
1406                         } else if (trylock_page(page)) {
1407                                 get_page(page);
1408                                 xas_unlock_irq(&xas);
1409                         } else {
1410                                 result = SCAN_PAGE_LOCK;
1411                                 goto xa_locked;
1412                         }
1413                 } else {        /* !is_shmem */
1414                         if (!page || xa_is_value(page)) {
1415                                 xas_unlock_irq(&xas);
1416                                 page_cache_sync_readahead(mapping, &file->f_ra,
1417                                                           file, index,
1418                                                           PAGE_SIZE);
1419                                 /* drain pagevecs to help isolate_lru_page() */
1420                                 lru_add_drain();
1421                                 page = find_lock_page(mapping, index);
1422                                 if (unlikely(page == NULL)) {
1423                                         result = SCAN_FAIL;
1424                                         goto xa_unlocked;
1425                                 }
1426                         } else if (!PageUptodate(page)) {
1427                                 xas_unlock_irq(&xas);
1428                                 wait_on_page_locked(page);
1429                                 if (!trylock_page(page)) {
1430                                         result = SCAN_PAGE_LOCK;
1431                                         goto xa_unlocked;
1432                                 }
1433                                 get_page(page);
1434                         } else if (PageDirty(page)) {
1435                                 result = SCAN_FAIL;
1436                                 goto xa_locked;
1437                         } else if (trylock_page(page)) {
1438                                 get_page(page);
1439                                 xas_unlock_irq(&xas);
1440                         } else {
1441                                 result = SCAN_PAGE_LOCK;
1442                                 goto xa_locked;
1443                         }
1444                 }
1445
1446                 /*
1447                  * The page must be locked, so we can drop the i_pages lock
1448                  * without racing with truncate.
1449                  */
1450                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1451                 VM_BUG_ON_PAGE(!PageUptodate(page), page);
1452
1453                 /*
1454                  * If file was truncated then extended, or hole-punched, before
1455                  * we locked the first page, then a THP might be there already.
1456                  */
1457                 if (PageTransCompound(page)) {
1458                         result = SCAN_PAGE_COMPOUND;
1459                         goto out_unlock;
1460                 }
1461
1462                 if (page_mapping(page) != mapping) {
1463                         result = SCAN_TRUNCATED;
1464                         goto out_unlock;
1465                 }
1466
1467                 if (isolate_lru_page(page)) {
1468                         result = SCAN_DEL_PAGE_LRU;
1469                         goto out_unlock;
1470                 }
1471
1472                 if (page_has_private(page) &&
1473                     !try_to_release_page(page, GFP_KERNEL)) {
1474                         result = SCAN_PAGE_HAS_PRIVATE;
1475                         goto out_unlock;
1476                 }
1477
1478                 if (page_mapped(page))
1479                         unmap_mapping_pages(mapping, index, 1, false);
1480
1481                 xas_lock_irq(&xas);
1482                 xas_set(&xas, index);
1483
1484                 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1485                 VM_BUG_ON_PAGE(page_mapped(page), page);
1486
1487                 /*
1488                  * The page is expected to have page_count() == 3:
1489                  *  - we hold a pin on it;
1490                  *  - one reference from page cache;
1491                  *  - one from isolate_lru_page;
1492                  */
1493                 if (!page_ref_freeze(page, 3)) {
1494                         result = SCAN_PAGE_COUNT;
1495                         xas_unlock_irq(&xas);
1496                         putback_lru_page(page);
1497                         goto out_unlock;
1498                 }
1499
1500                 /*
1501                  * Add the page to the list to be able to undo the collapse if
1502                  * something go wrong.
1503                  */
1504                 list_add_tail(&page->lru, &pagelist);
1505
1506                 /* Finally, replace with the new page. */
1507                 xas_store(&xas, new_page);
1508                 continue;
1509 out_unlock:
1510                 unlock_page(page);
1511                 put_page(page);
1512                 goto xa_unlocked;
1513         }
1514
1515         if (is_shmem)
1516                 __inc_node_page_state(new_page, NR_SHMEM_THPS);
1517         else {
1518                 __inc_node_page_state(new_page, NR_FILE_THPS);
1519                 filemap_nr_thps_inc(mapping);
1520         }
1521
1522         if (nr_none) {
1523                 struct zone *zone = page_zone(new_page);
1524
1525                 __mod_node_page_state(zone->zone_pgdat, NR_FILE_PAGES, nr_none);
1526                 if (is_shmem)
1527                         __mod_node_page_state(zone->zone_pgdat,
1528                                               NR_SHMEM, nr_none);
1529         }
1530
1531 xa_locked:
1532         xas_unlock_irq(&xas);
1533 xa_unlocked:
1534
1535         if (result == SCAN_SUCCEED) {
1536                 struct page *page, *tmp;
1537
1538                 /*
1539                  * Replacing old pages with new one has succeeded, now we
1540                  * need to copy the content and free the old pages.
1541                  */
1542                 index = start;
1543                 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1544                         while (index < page->index) {
1545                                 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1546                                 index++;
1547                         }
1548                         copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1549                                         page);
1550                         list_del(&page->lru);
1551                         page->mapping = NULL;
1552                         page_ref_unfreeze(page, 1);
1553                         ClearPageActive(page);
1554                         ClearPageUnevictable(page);
1555                         unlock_page(page);
1556                         put_page(page);
1557                         index++;
1558                 }
1559                 while (index < end) {
1560                         clear_highpage(new_page + (index % HPAGE_PMD_NR));
1561                         index++;
1562                 }
1563
1564                 SetPageUptodate(new_page);
1565                 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1566                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1567
1568                 if (is_shmem) {
1569                         set_page_dirty(new_page);
1570                         lru_cache_add_anon(new_page);
1571                 } else {
1572                         lru_cache_add_file(new_page);
1573                 }
1574                 count_memcg_events(memcg, THP_COLLAPSE_ALLOC, 1);
1575
1576                 /*
1577                  * Remove pte page tables, so we can re-fault the page as huge.
1578                  */
1579                 retract_page_tables(mapping, start);
1580                 *hpage = NULL;
1581
1582                 khugepaged_pages_collapsed++;
1583         } else {
1584                 struct page *page;
1585
1586                 /* Something went wrong: roll back page cache changes */
1587                 xas_lock_irq(&xas);
1588                 mapping->nrpages -= nr_none;
1589
1590                 if (is_shmem)
1591                         shmem_uncharge(mapping->host, nr_none);
1592
1593                 xas_set(&xas, start);
1594                 xas_for_each(&xas, page, end - 1) {
1595                         page = list_first_entry_or_null(&pagelist,
1596                                         struct page, lru);
1597                         if (!page || xas.xa_index < page->index) {
1598                                 if (!nr_none)
1599                                         break;
1600                                 nr_none--;
1601                                 /* Put holes back where they were */
1602                                 xas_store(&xas, NULL);
1603                                 continue;
1604                         }
1605
1606                         VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1607
1608                         /* Unfreeze the page. */
1609                         list_del(&page->lru);
1610                         page_ref_unfreeze(page, 2);
1611                         xas_store(&xas, page);
1612                         xas_pause(&xas);
1613                         xas_unlock_irq(&xas);
1614                         unlock_page(page);
1615                         putback_lru_page(page);
1616                         xas_lock_irq(&xas);
1617                 }
1618                 VM_BUG_ON(nr_none);
1619                 xas_unlock_irq(&xas);
1620
1621                 mem_cgroup_cancel_charge(new_page, memcg, true);
1622                 new_page->mapping = NULL;
1623         }
1624
1625         unlock_page(new_page);
1626 out:
1627         VM_BUG_ON(!list_empty(&pagelist));
1628         /* TODO: tracepoints */
1629 }
1630
1631 static void khugepaged_scan_file(struct mm_struct *mm,
1632                 struct file *file, pgoff_t start, struct page **hpage)
1633 {
1634         struct page *page = NULL;
1635         struct address_space *mapping = file->f_mapping;
1636         XA_STATE(xas, &mapping->i_pages, start);
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         xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
1646                 if (xas_retry(&xas, page))
1647                         continue;
1648
1649                 if (xa_is_value(page)) {
1650                         if (++swap > khugepaged_max_ptes_swap) {
1651                                 result = SCAN_EXCEED_SWAP_PTE;
1652                                 break;
1653                         }
1654                         continue;
1655                 }
1656
1657                 if (PageTransCompound(page)) {
1658                         result = SCAN_PAGE_COMPOUND;
1659                         break;
1660                 }
1661
1662                 node = page_to_nid(page);
1663                 if (khugepaged_scan_abort(node)) {
1664                         result = SCAN_SCAN_ABORT;
1665                         break;
1666                 }
1667                 khugepaged_node_load[node]++;
1668
1669                 if (!PageLRU(page)) {
1670                         result = SCAN_PAGE_LRU;
1671                         break;
1672                 }
1673
1674                 if (page_count(page) !=
1675                     1 + page_mapcount(page) + page_has_private(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                         xas_pause(&xas);
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_file(mm, file, start, hpage, node);
1701                 }
1702         }
1703
1704         /* TODO: tracepoints */
1705 }
1706 #else
1707 static void khugepaged_scan_file(struct mm_struct *mm,
1708                 struct file *file, pgoff_t start, struct page **hpage)
1709 {
1710         BUILD_BUG();
1711 }
1712 #endif
1713
1714 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
1715                                             struct page **hpage)
1716         __releases(&khugepaged_mm_lock)
1717         __acquires(&khugepaged_mm_lock)
1718 {
1719         struct mm_slot *mm_slot;
1720         struct mm_struct *mm;
1721         struct vm_area_struct *vma;
1722         int progress = 0;
1723
1724         VM_BUG_ON(!pages);
1725         lockdep_assert_held(&khugepaged_mm_lock);
1726
1727         if (khugepaged_scan.mm_slot)
1728                 mm_slot = khugepaged_scan.mm_slot;
1729         else {
1730                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
1731                                      struct mm_slot, mm_node);
1732                 khugepaged_scan.address = 0;
1733                 khugepaged_scan.mm_slot = mm_slot;
1734         }
1735         spin_unlock(&khugepaged_mm_lock);
1736
1737         mm = mm_slot->mm;
1738         /*
1739          * Don't wait for semaphore (to avoid long wait times).  Just move to
1740          * the next mm on the list.
1741          */
1742         vma = NULL;
1743         if (unlikely(!down_read_trylock(&mm->mmap_sem)))
1744                 goto breakouterloop_mmap_sem;
1745         if (likely(!khugepaged_test_exit(mm)))
1746                 vma = find_vma(mm, khugepaged_scan.address);
1747
1748         progress++;
1749         for (; vma; vma = vma->vm_next) {
1750                 unsigned long hstart, hend;
1751
1752                 cond_resched();
1753                 if (unlikely(khugepaged_test_exit(mm))) {
1754                         progress++;
1755                         break;
1756                 }
1757                 if (!hugepage_vma_check(vma, vma->vm_flags)) {
1758 skip:
1759                         progress++;
1760                         continue;
1761                 }
1762                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1763                 hend = vma->vm_end & HPAGE_PMD_MASK;
1764                 if (hstart >= hend)
1765                         goto skip;
1766                 if (khugepaged_scan.address > hend)
1767                         goto skip;
1768                 if (khugepaged_scan.address < hstart)
1769                         khugepaged_scan.address = hstart;
1770                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
1771
1772                 while (khugepaged_scan.address < hend) {
1773                         int ret;
1774                         cond_resched();
1775                         if (unlikely(khugepaged_test_exit(mm)))
1776                                 goto breakouterloop;
1777
1778                         VM_BUG_ON(khugepaged_scan.address < hstart ||
1779                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
1780                                   hend);
1781                         if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
1782                                 struct file *file;
1783                                 pgoff_t pgoff = linear_page_index(vma,
1784                                                 khugepaged_scan.address);
1785
1786                                 if (shmem_file(vma->vm_file)
1787                                     && !shmem_huge_enabled(vma))
1788                                         goto skip;
1789                                 file = get_file(vma->vm_file);
1790                                 up_read(&mm->mmap_sem);
1791                                 ret = 1;
1792                                 khugepaged_scan_file(mm, file, 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         static struct task_struct *khugepaged_thread __read_mostly;
1981         static DEFINE_MUTEX(khugepaged_mutex);
1982         int err = 0;
1983
1984         mutex_lock(&khugepaged_mutex);
1985         if (khugepaged_enabled()) {
1986                 if (!khugepaged_thread)
1987                         khugepaged_thread = kthread_run(khugepaged, NULL,
1988                                                         "khugepaged");
1989                 if (IS_ERR(khugepaged_thread)) {
1990                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
1991                         err = PTR_ERR(khugepaged_thread);
1992                         khugepaged_thread = NULL;
1993                         goto fail;
1994                 }
1995
1996                 if (!list_empty(&khugepaged_scan.mm_head))
1997                         wake_up_interruptible(&khugepaged_wait);
1998
1999                 set_recommended_min_free_kbytes();
2000         } else if (khugepaged_thread) {
2001                 kthread_stop(khugepaged_thread);
2002                 khugepaged_thread = NULL;
2003         }
2004 fail:
2005         mutex_unlock(&khugepaged_mutex);
2006         return err;
2007 }