mm/khugepaged.c

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