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