mm/khugepaged.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   3
   4 #include <linux/mm.h>
   5 #include <linux/sched.h>
   6 #include <linux/sched/mm.h>
   7 #include <linux/sched/coredump.h>
   8 #include <linux/mmu_notifier.h>
   9 #include <linux/rmap.h>
  10 #include <linux/swap.h>
  11 #include <linux/mm_inline.h>
  12 #include <linux/kthread.h>
  13 #include <linux/khugepaged.h>
  14 #include <linux/freezer.h>
  15 #include <linux/mman.h>
  16 #include <linux/hashtable.h>
  17 #include <linux/userfaultfd_k.h>
  18 #include <linux/page_idle.h>
  19 #include <linux/swapops.h>
  20 #include <linux/shmem_fs.h>
  21
  22 #include <asm/tlb.h>
  23 #include <asm/pgalloc.h>
  24 #include "internal.h"
  25
  26 enum scan_result {
  27         SCAN_FAIL,
  28         SCAN_SUCCEED,
  29         SCAN_PMD_NULL,
  30         SCAN_EXCEED_NONE_PTE,
  31         SCAN_PTE_NON_PRESENT,
  32         SCAN_PAGE_RO,
  33         SCAN_LACK_REFERENCED_PAGE,
  34         SCAN_PAGE_NULL,
  35         SCAN_SCAN_ABORT,
  36         SCAN_PAGE_COUNT,
  37         SCAN_PAGE_LRU,
  38         SCAN_PAGE_LOCK,
  39         SCAN_PAGE_ANON,
  40         SCAN_PAGE_COMPOUND,
  41         SCAN_ANY_PROCESS,
  42         SCAN_VMA_NULL,
  43         SCAN_VMA_CHECK,
  44         SCAN_ADDRESS_RANGE,
  45         SCAN_SWAP_CACHE_PAGE,
  46         SCAN_DEL_PAGE_LRU,
  47         SCAN_ALLOC_HUGE_PAGE_FAIL,
  48         SCAN_CGROUP_CHARGE_FAIL,
  49         SCAN_EXCEED_SWAP_PTE,
  50         SCAN_TRUNCATED,
  51         SCAN_PAGE_HAS_PRIVATE,
  52 };
  53
  54 #define CREATE_TRACE_POINTS
  55 #include <trace/events/huge_memory.h>
  56
  57 /* default scan 8*512 pte (or vmas) every 30 second */
  58 static unsigned int khugepaged_pages_to_scan __read_mostly;
  59 static unsigned int khugepaged_pages_collapsed;
  60 static unsigned int khugepaged_full_scans;
  61 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  62 /* during fragmentation poll the hugepage allocator once every minute */
  63 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  64 static unsigned long khugepaged_sleep_expire;
  65 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  66 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  67 /*
  68  * default collapse hugepages if there is at least one pte mapped like
  69  * it would have happened if the vma was large enough during page
  70  * fault.
  71  */
  72 static unsigned int khugepaged_max_ptes_none __read_mostly;
  73 static unsigned int khugepaged_max_ptes_swap __read_mostly;
  74
  75 #define MM_SLOTS_HASH_BITS 10
  76 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  77
  78 static struct kmem_cache *mm_slot_cache __read_mostly;
  79
  80 #define MAX_PTE_MAPPED_THP 8
  81
  82 /**
  83  * struct mm_slot - hash lookup from mm to mm_slot
  84  * @hash: hash collision list
  85  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  86  * @mm: the mm that this information is valid for
  87  */
  88 struct mm_slot {
  89         struct hlist_node hash;
  90         struct list_head mm_node;
  91         struct mm_struct *mm;
  92
  93         /* pte-mapped THP in this mm */
  94         int nr_pte_mapped_thp;
  95         unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP];
  96 };
  97
  98 /**
  99  * struct khugepaged_scan - cursor for scanning
 100  * @mm_head: the head of the mm list to scan
 101  * @mm_slot: the current mm_slot we are scanning
 102  * @address: the next address inside that to be scanned
 103  *
 104  * There is only the one khugepaged_scan instance of this cursor structure.
 105  */
 106 struct khugepaged_scan {
 107         struct list_head mm_head;
 108         struct mm_slot *mm_slot;
 109         unsigned long address;
 110 };
 111
 112 static struct khugepaged_scan khugepaged_scan = {
 113         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 114 };
 115
 116 #ifdef CONFIG_SYSFS
 117 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 118                                          struct kobj_attribute *attr,
 119                                          char *buf)
 120 {
 121         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 122 }
 123
 124 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 125                                           struct kobj_attribute *attr,
 126                                           const char *buf, size_t count)
 127 {
 128         unsigned long msecs;
 129         int err;
 130
 131         err = kstrtoul(buf, 10, &msecs);
 132         if (err || msecs > UINT_MAX)
 133                 return -EINVAL;
 134
 135         khugepaged_scan_sleep_millisecs = msecs;
 136         khugepaged_sleep_expire = 0;
 137         wake_up_interruptible(&khugepaged_wait);
 138
 139         return count;
 140 }
 141 static struct kobj_attribute scan_sleep_millisecs_attr =
 142         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 143                scan_sleep_millisecs_store);
 144
 145 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 146                                           struct kobj_attribute *attr,
 147                                           char *buf)
 148 {
 149         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 150 }
 151
 152 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 153                                            struct kobj_attribute *attr,
 154                                            const char *buf, size_t count)
 155 {
 156         unsigned long msecs;
 157         int err;
 158
 159         err = kstrtoul(buf, 10, &msecs);
 160         if (err || msecs > UINT_MAX)
 161                 return -EINVAL;
 162
 163         khugepaged_alloc_sleep_millisecs = msecs;
 164         khugepaged_sleep_expire = 0;
 165         wake_up_interruptible(&khugepaged_wait);
 166
 167         return count;
 168 }
 169 static struct kobj_attribute alloc_sleep_millisecs_attr =
 170         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 171                alloc_sleep_millisecs_store);
 172
 173 static ssize_t pages_to_scan_show(struct kobject *kobj,
 174                                   struct kobj_attribute *attr,
 175                                   char *buf)
 176 {
 177         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 178 }
 179 static ssize_t pages_to_scan_store(struct kobject *kobj,
 180                                    struct kobj_attribute *attr,
 181                                    const char *buf, size_t count)
 182 {
 183         int err;
 184         unsigned long pages;
 185
 186         err = kstrtoul(buf, 10, &pages);
 187         if (err || !pages || pages > UINT_MAX)
 188                 return -EINVAL;
 189
 190         khugepaged_pages_to_scan = pages;
 191
 192         return count;
 193 }
 194 static struct kobj_attribute pages_to_scan_attr =
 195         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 196                pages_to_scan_store);
 197
 198 static ssize_t pages_collapsed_show(struct kobject *kobj,
 199                                     struct kobj_attribute *attr,
 200                                     char *buf)
 201 {
 202         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 203 }
 204 static struct kobj_attribute pages_collapsed_attr =
 205         __ATTR_RO(pages_collapsed);
 206
 207 static ssize_t full_scans_show(struct kobject *kobj,
 208                                struct kobj_attribute *attr,
 209                                char *buf)
 210 {
 211         return sprintf(buf, "%u\n", khugepaged_full_scans);
 212 }
 213 static struct kobj_attribute full_scans_attr =
 214         __ATTR_RO(full_scans);
 215
 216 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 217                                       struct kobj_attribute *attr, char *buf)
 218 {
 219         return single_hugepage_flag_show(kobj, attr, buf,
 220                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 221 }
 222 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 223                                        struct kobj_attribute *attr,
 224                                        const char *buf, size_t count)
 225 {
 226         return single_hugepage_flag_store(kobj, attr, buf, count,
 227                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 228 }
 229 static struct kobj_attribute khugepaged_defrag_attr =
 230         __ATTR(defrag, 0644, khugepaged_defrag_show,
 231                khugepaged_defrag_store);
 232
 233 /*
 234  * max_ptes_none controls if khugepaged should collapse hugepages over
 235  * any unmapped ptes in turn potentially increasing the memory
 236  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 237  * reduce the available free memory in the system as it
 238  * runs. Increasing max_ptes_none will instead potentially reduce the
 239  * free memory in the system during the khugepaged scan.
 240  */
 241 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 242                                              struct kobj_attribute *attr,
 243                                              char *buf)
 244 {
 245         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 246 }
 247 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 248                                               struct kobj_attribute *attr,
 249                                               const char *buf, size_t count)
 250 {
 251         int err;
 252         unsigned long max_ptes_none;
 253
 254         err = kstrtoul(buf, 10, &max_ptes_none);
 255         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 256                 return -EINVAL;
 257
 258         khugepaged_max_ptes_none = max_ptes_none;
 259
 260         return count;
 261 }
 262 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 263         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 264                khugepaged_max_ptes_none_store);
 265
 266 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
 267                                              struct kobj_attribute *attr,
 268                                              char *buf)
 269 {
 270         return sprintf(buf, "%u\n", khugepaged_max_ptes_swap);
 271 }
 272
 273 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
 274                                               struct kobj_attribute *attr,
 275                                               const char *buf, size_t count)
 276 {
 277         int err;
 278         unsigned long max_ptes_swap;
 279
 280         err  = kstrtoul(buf, 10, &max_ptes_swap);
 281         if (err || max_ptes_swap > HPAGE_PMD_NR-1)
 282                 return -EINVAL;
 283
 284         khugepaged_max_ptes_swap = max_ptes_swap;
 285
 286         return count;
 287 }
 288
 289 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
 290         __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
 291                khugepaged_max_ptes_swap_store);
 292
 293 static struct attribute *khugepaged_attr[] = {
 294         &khugepaged_defrag_attr.attr,
 295         &khugepaged_max_ptes_none_attr.attr,
 296         &pages_to_scan_attr.attr,
 297         &pages_collapsed_attr.attr,
 298         &full_scans_attr.attr,
 299         &scan_sleep_millisecs_attr.attr,
 300         &alloc_sleep_millisecs_attr.attr,
 301         &khugepaged_max_ptes_swap_attr.attr,
 302         NULL,
 303 };
 304
 305 struct attribute_group khugepaged_attr_group = {
 306         .attrs = khugepaged_attr,
 307         .name = "khugepaged",
 308 };
 309 #endif /* CONFIG_SYSFS */
 310
 311 #define VM_NO_KHUGEPAGED (VM_SPECIAL | VM_HUGETLB)
 312
 313 int hugepage_madvise(struct vm_area_struct *vma,
 314                      unsigned long *vm_flags, int advice)
 315 {
 316         switch (advice) {
 317         case MADV_HUGEPAGE:
 318 #ifdef CONFIG_S390
 319                 /*
 320                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
 321                  * can't handle this properly after s390_enable_sie, so we simply
 322                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
 323                  */
 324                 if (mm_has_pgste(vma->vm_mm))
 325                         return 0;
 326 #endif
 327                 *vm_flags &= ~VM_NOHUGEPAGE;
 328                 *vm_flags |= VM_HUGEPAGE;
 329                 /*
 330                  * If the vma become good for khugepaged to scan,
 331                  * register it here without waiting a page fault that
 332                  * may not happen any time soon.
 333                  */
 334                 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
 335                                 khugepaged_enter_vma_merge(vma, *vm_flags))
 336                         return -ENOMEM;
 337                 break;
 338         case MADV_NOHUGEPAGE:
 339                 *vm_flags &= ~VM_HUGEPAGE;
 340                 *vm_flags |= VM_NOHUGEPAGE;
 341                 /*
 342                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
 343                  * this vma even if we leave the mm registered in khugepaged if
 344                  * it got registered before VM_NOHUGEPAGE was set.
 345                  */
 346                 break;
 347         }
 348
 349         return 0;
 350 }
 351
 352 int __init khugepaged_init(void)
 353 {
 354         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
 355                                           sizeof(struct mm_slot),
 356                                           __alignof__(struct mm_slot), 0, NULL);
 357         if (!mm_slot_cache)
 358                 return -ENOMEM;
 359
 360         khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
 361         khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
 362         khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
 363
 364         return 0;
 365 }
 366
 367 void __init khugepaged_destroy(void)
 368 {
 369         kmem_cache_destroy(mm_slot_cache);
 370 }
 371
 372 static inline struct mm_slot *alloc_mm_slot(void)
 373 {
 374         if (!mm_slot_cache)     /* initialization failed */
 375                 return NULL;
 376         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
 377 }
 378
 379 static inline void free_mm_slot(struct mm_slot *mm_slot)
 380 {
 381         kmem_cache_free(mm_slot_cache, mm_slot);
 382 }
 383
 384 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
 385 {
 386         struct mm_slot *mm_slot;
 387
 388         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
 389                 if (mm == mm_slot->mm)
 390                         return mm_slot;
 391
 392         return NULL;
 393 }
 394
 395 static void insert_to_mm_slots_hash(struct mm_struct *mm,
 396                                     struct mm_slot *mm_slot)
 397 {
 398         mm_slot->mm = mm;
 399         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
 400 }
 401
 402 static inline int khugepaged_test_exit(struct mm_struct *mm)
 403 {
 404         return atomic_read(&mm->mm_users) == 0;
 405 }
 406
 407 static bool hugepage_vma_check(struct vm_area_struct *vma,
 408                                unsigned long vm_flags)
 409 {
 410         if ((!(vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
 411             (vm_flags & VM_NOHUGEPAGE) ||
 412             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
 413                 return false;
 414
 415         if (shmem_file(vma->vm_file) ||
 416             (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) &&
 417              vma->vm_file &&
 418              (vm_flags & VM_DENYWRITE))) {
 419                 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
 420                         return false;
 421                 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
 422                                 HPAGE_PMD_NR);
 423         }
 424         if (!vma->anon_vma || vma->vm_ops)
 425                 return false;
 426         if (is_vma_temporary_stack(vma))
 427                 return false;
 428         return !(vm_flags & VM_NO_KHUGEPAGED);
 429 }
 430
 431 int __khugepaged_enter(struct mm_struct *mm)
 432 {
 433         struct mm_slot *mm_slot;
 434         int wakeup;
 435
 436         mm_slot = alloc_mm_slot();
 437         if (!mm_slot)
 438                 return -ENOMEM;
 439
 440         /* __khugepaged_exit() must not run from under us */
 441         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
 442         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
 443                 free_mm_slot(mm_slot);
 444                 return 0;
 445         }
 446
 447         spin_lock(&khugepaged_mm_lock);
 448         insert_to_mm_slots_hash(mm, mm_slot);
 449         /*
 450          * Insert just behind the scanning cursor, to let the area settle
 451          * down a little.
 452          */
 453         wakeup = list_empty(&khugepaged_scan.mm_head);
 454         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
 455         spin_unlock(&khugepaged_mm_lock);
 456
 457         mmgrab(mm);
 458         if (wakeup)
 459                 wake_up_interruptible(&khugepaged_wait);
 460
 461         return 0;
 462 }
 463
 464 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
 465                                unsigned long vm_flags)
 466 {
 467         unsigned long hstart, hend;
 468
 469         /*
 470          * khugepaged only supports read-only files for non-shmem files.
 471          * khugepaged does not yet work on special mappings. And
 472          * file-private shmem THP is not supported.
 473          */
 474         if (!hugepage_vma_check(vma, vm_flags))
 475                 return 0;
 476
 477         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 478         hend = vma->vm_end & HPAGE_PMD_MASK;
 479         if (hstart < hend)
 480                 return khugepaged_enter(vma, vm_flags);
 481         return 0;
 482 }
 483
 484 void __khugepaged_exit(struct mm_struct *mm)
 485 {
 486         struct mm_slot *mm_slot;
 487         int free = 0;
 488
 489         spin_lock(&khugepaged_mm_lock);
 490         mm_slot = get_mm_slot(mm);
 491         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
 492                 hash_del(&mm_slot->hash);
 493                 list_del(&mm_slot->mm_node);
 494                 free = 1;
 495         }
 496         spin_unlock(&khugepaged_mm_lock);
 497
 498         if (free) {
 499                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
 500                 free_mm_slot(mm_slot);
 501                 mmdrop(mm);
 502         } else if (mm_slot) {
 503                 /*
 504                  * This is required to serialize against
 505                  * khugepaged_test_exit() (which is guaranteed to run
 506                  * under mmap sem read mode). Stop here (after we
 507                  * return all pagetables will be destroyed) until
 508                  * khugepaged has finished working on the pagetables
 509                  * under the mmap_sem.
 510                  */
 511                 down_write(&mm->mmap_sem);
 512                 up_write(&mm->mmap_sem);
 513         }
 514 }
 515
 516 static void release_pte_page(struct page *page)
 517 {
 518         dec_node_page_state(page, NR_ISOLATED_ANON + page_is_file_cache(page));
 519         unlock_page(page);
 520         putback_lru_page(page);
 521 }
 522
 523 static void release_pte_pages(pte_t *pte, pte_t *_pte)
 524 {
 525         while (--_pte >= pte) {
 526                 pte_t pteval = *_pte;
 527                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
 528                         release_pte_page(pte_page(pteval));
 529         }
 530 }
 531
 532 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
 533                                         unsigned long address,
 534                                         pte_t *pte)
 535 {
 536         struct page *page = NULL;
 537         pte_t *_pte;
 538         int none_or_zero = 0, result = 0, referenced = 0;
 539         bool writable = false;
 540
 541         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
 542              _pte++, address += PAGE_SIZE) {
 543                 pte_t pteval = *_pte;
 544                 if (pte_none(pteval) || (pte_present(pteval) &&
 545                                 is_zero_pfn(pte_pfn(pteval)))) {
 546                         if (!userfaultfd_armed(vma) &&
 547                             ++none_or_zero <= khugepaged_max_ptes_none) {
 548                                 continue;
 549                         } else {
 550                                 result = SCAN_EXCEED_NONE_PTE;
 551                                 goto out;
 552                         }
 553                 }
 554                 if (!pte_present(pteval)) {
 555                         result = SCAN_PTE_NON_PRESENT;
 556                         goto out;
 557                 }
 558                 page = vm_normal_page(vma, address, pteval);
 559                 if (unlikely(!page)) {
 560                         result = SCAN_PAGE_NULL;
 561                         goto out;
 562                 }
 563
 564                 /* TODO: teach khugepaged to collapse THP mapped with pte */
 565                 if (PageCompound(page)) {
 566                         result = SCAN_PAGE_COMPOUND;
 567                         goto out;
 568                 }
 569
 570                 VM_BUG_ON_PAGE(!PageAnon(page), page);
 571
 572                 /*
 573                  * We can do it before isolate_lru_page because the
 574                  * page can't be freed from under us. NOTE: PG_lock
 575                  * is needed to serialize against split_huge_page
 576                  * when invoked from the VM.
 577                  */
 578                 if (!trylock_page(page)) {
 579                         result = SCAN_PAGE_LOCK;
 580                         goto out;
 581                 }
 582
 583                 /*
 584                  * cannot use mapcount: can't collapse if there's a gup pin.
 585                  * The page must only be referenced by the scanned process
 586                  * and page swap cache.
 587                  */
 588                 if (page_count(page) != 1 + PageSwapCache(page)) {
 589                         unlock_page(page);
 590                         result = SCAN_PAGE_COUNT;
 591                         goto out;
 592                 }
 593                 if (pte_write(pteval)) {
 594                         writable = true;
 595                 } else {
 596                         if (PageSwapCache(page) &&
 597                             !reuse_swap_page(page, NULL)) {
 598                                 unlock_page(page);
 599                                 result = SCAN_SWAP_CACHE_PAGE;
 600                                 goto out;
 601                         }
 602                         /*
 603                          * Page is not in the swap cache. It can be collapsed
 604                          * into a THP.
 605                          */
 606                 }
 607
 608                 /*
 609                  * Isolate the page to avoid collapsing an hugepage
 610                  * currently in use by the VM.
 611                  */
 612                 if (isolate_lru_page(page)) {
 613                         unlock_page(page);
 614                         result = SCAN_DEL_PAGE_LRU;
 615                         goto out;
 616                 }
 617                 inc_node_page_state(page,
 618                                 NR_ISOLATED_ANON + page_is_file_cache(page));
 619                 VM_BUG_ON_PAGE(!PageLocked(page), page);
 620                 VM_BUG_ON_PAGE(PageLRU(page), page);
 621
 622                 /* There should be enough young pte to collapse the page */
 623                 if (pte_young(pteval) ||
 624                     page_is_young(page) || PageReferenced(page) ||
 625                     mmu_notifier_test_young(vma->vm_mm, address))
 626                         referenced++;
 627         }
 628         if (likely(writable)) {
 629                 if (likely(referenced)) {
 630                         result = SCAN_SUCCEED;
 631                         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 632                                                             referenced, writable, result);
 633                         return 1;
 634                 }
 635         } else {
 636                 result = SCAN_PAGE_RO;
 637         }
 638
 639 out:
 640         release_pte_pages(pte, _pte);
 641         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 642                                             referenced, writable, result);
 643         return 0;
 644 }
 645
 646 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
 647                                       struct vm_area_struct *vma,
 648                                       unsigned long address,
 649                                       spinlock_t *ptl)
 650 {
 651         pte_t *_pte;
 652         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 653                                 _pte++, page++, address += PAGE_SIZE) {
 654                 pte_t pteval = *_pte;
 655                 struct page *src_page;
 656
 657                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
 658                         clear_user_highpage(page, address);
 659                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
 660                         if (is_zero_pfn(pte_pfn(pteval))) {
 661                                 /*
 662                                  * ptl mostly unnecessary.
 663                                  */
 664                                 spin_lock(ptl);
 665                                 /*
 666                                  * paravirt calls inside pte_clear here are
 667                                  * superfluous.
 668                                  */
 669                                 pte_clear(vma->vm_mm, address, _pte);
 670                                 spin_unlock(ptl);
 671                         }
 672                 } else {
 673                         src_page = pte_page(pteval);
 674                         copy_user_highpage(page, src_page, address, vma);
 675                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
 676                         release_pte_page(src_page);
 677                         /*
 678                          * ptl mostly unnecessary, but preempt has to
 679                          * be disabled to update the per-cpu stats
 680                          * inside page_remove_rmap().
 681                          */
 682                         spin_lock(ptl);
 683                         /*
 684                          * paravirt calls inside pte_clear here are
 685                          * superfluous.
 686                          */
 687                         pte_clear(vma->vm_mm, address, _pte);
 688                         page_remove_rmap(src_page, false);
 689                         spin_unlock(ptl);
 690                         free_page_and_swap_cache(src_page);
 691                 }
 692         }
 693 }
 694
 695 static void khugepaged_alloc_sleep(void)
 696 {
 697         DEFINE_WAIT(wait);
 698
 699         add_wait_queue(&khugepaged_wait, &wait);
 700         freezable_schedule_timeout_interruptible(
 701                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
 702         remove_wait_queue(&khugepaged_wait, &wait);
 703 }
 704
 705 static int khugepaged_node_load[MAX_NUMNODES];
 706
 707 static bool khugepaged_scan_abort(int nid)
 708 {
 709         int i;
 710
 711         /*
 712          * If node_reclaim_mode is disabled, then no extra effort is made to
 713          * allocate memory locally.
 714          */
 715         if (!node_reclaim_mode)
 716                 return false;
 717
 718         /* If there is a count for this node already, it must be acceptable */
 719         if (khugepaged_node_load[nid])
 720                 return false;
 721
 722         for (i = 0; i < MAX_NUMNODES; i++) {
 723                 if (!khugepaged_node_load[i])
 724                         continue;
 725                 if (node_distance(nid, i) > node_reclaim_distance)
 726                         return true;
 727         }
 728         return false;
 729 }
 730
 731 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
 732 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
 733 {
 734         return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
 735 }
 736
 737 #ifdef CONFIG_NUMA
 738 static int khugepaged_find_target_node(void)
 739 {
 740         static int last_khugepaged_target_node = NUMA_NO_NODE;
 741         int nid, target_node = 0, max_value = 0;
 742
 743         /* find first node with max normal pages hit */
 744         for (nid = 0; nid < MAX_NUMNODES; nid++)
 745                 if (khugepaged_node_load[nid] > max_value) {
 746                         max_value = khugepaged_node_load[nid];
 747                         target_node = nid;
 748                 }
 749
 750         /* do some balance if several nodes have the same hit record */
 751         if (target_node <= last_khugepaged_target_node)
 752                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
 753                                 nid++)
 754                         if (max_value == khugepaged_node_load[nid]) {
 755                                 target_node = nid;
 756                                 break;
 757                         }
 758
 759         last_khugepaged_target_node = target_node;
 760         return target_node;
 761 }
 762
 763 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 764 {
 765         if (IS_ERR(*hpage)) {
 766                 if (!*wait)
 767                         return false;
 768
 769                 *wait = false;
 770                 *hpage = NULL;
 771                 khugepaged_alloc_sleep();
 772         } else if (*hpage) {
 773                 put_page(*hpage);
 774                 *hpage = NULL;
 775         }
 776
 777         return true;
 778 }
 779
 780 static struct page *
 781 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 782 {
 783         VM_BUG_ON_PAGE(*hpage, *hpage);
 784
 785         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
 786         if (unlikely(!*hpage)) {
 787                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 788                 *hpage = ERR_PTR(-ENOMEM);
 789                 return NULL;
 790         }
 791
 792         prep_transhuge_page(*hpage);
 793         count_vm_event(THP_COLLAPSE_ALLOC);
 794         return *hpage;
 795 }
 796 #else
 797 static int khugepaged_find_target_node(void)
 798 {
 799         return 0;
 800 }
 801
 802 static inline struct page *alloc_khugepaged_hugepage(void)
 803 {
 804         struct page *page;
 805
 806         page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
 807                            HPAGE_PMD_ORDER);
 808         if (page)
 809                 prep_transhuge_page(page);
 810         return page;
 811 }
 812
 813 static struct page *khugepaged_alloc_hugepage(bool *wait)
 814 {
 815         struct page *hpage;
 816
 817         do {
 818                 hpage = alloc_khugepaged_hugepage();
 819                 if (!hpage) {
 820                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 821                         if (!*wait)
 822                                 return NULL;
 823
 824                         *wait = false;
 825                         khugepaged_alloc_sleep();
 826                 } else
 827                         count_vm_event(THP_COLLAPSE_ALLOC);
 828         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
 829
 830         return hpage;
 831 }
 832
 833 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 834 {
 835         if (!*hpage)
 836                 *hpage = khugepaged_alloc_hugepage(wait);
 837
 838         if (unlikely(!*hpage))
 839                 return false;
 840
 841         return true;
 842 }
 843
 844 static struct page *
 845 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 846 {
 847         VM_BUG_ON(!*hpage);
 848
 849         return  *hpage;
 850 }
 851 #endif
 852
 853 /*
 854  * If mmap_sem temporarily dropped, revalidate vma
 855  * before taking mmap_sem.
 856  * Return 0 if succeeds, otherwise return none-zero
 857  * value (scan code).
 858  */
 859
 860 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
 861                 struct vm_area_struct **vmap)
 862 {
 863         struct vm_area_struct *vma;
 864         unsigned long hstart, hend;
 865
 866         if (unlikely(khugepaged_test_exit(mm)))
 867                 return SCAN_ANY_PROCESS;
 868
 869         *vmap = vma = find_vma(mm, address);
 870         if (!vma)
 871                 return SCAN_VMA_NULL;
 872
 873         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 874         hend = vma->vm_end & HPAGE_PMD_MASK;
 875         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
 876                 return SCAN_ADDRESS_RANGE;
 877         if (!hugepage_vma_check(vma, vma->vm_flags))
 878                 return SCAN_VMA_CHECK;
 879         return 0;
 880 }
 881
 882 /*
 883  * Bring missing pages in from swap, to complete THP collapse.
 884  * Only done if khugepaged_scan_pmd believes it is worthwhile.
 885  *
 886  * Called and returns without pte mapped or spinlocks held,
 887  * but with mmap_sem held to protect against vma changes.
 888  */
 889
 890 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
 891                                         struct vm_area_struct *vma,
 892                                         unsigned long address, pmd_t *pmd,
 893                                         int referenced)
 894 {
 895         int swapped_in = 0;
 896         vm_fault_t ret = 0;
 897         struct vm_fault vmf = {
 898                 .vma = vma,
 899                 .address = address,
 900                 .flags = FAULT_FLAG_ALLOW_RETRY,
 901                 .pmd = pmd,
 902                 .pgoff = linear_page_index(vma, address),
 903         };
 904
 905         /* we only decide to swapin, if there is enough young ptes */
 906         if (referenced < HPAGE_PMD_NR/2) {
 907                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 908                 return false;
 909         }
 910         vmf.pte = pte_offset_map(pmd, address);
 911         for (; vmf.address < address + HPAGE_PMD_NR*PAGE_SIZE;
 912                         vmf.pte++, vmf.address += PAGE_SIZE) {
 913                 vmf.orig_pte = *vmf.pte;
 914                 if (!is_swap_pte(vmf.orig_pte))
 915                         continue;
 916                 swapped_in++;
 917                 ret = do_swap_page(&vmf);
 918
 919                 /* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */
 920                 if (ret & VM_FAULT_RETRY) {
 921                         down_read(&mm->mmap_sem);
 922                         if (hugepage_vma_revalidate(mm, address, &vmf.vma)) {
 923                                 /* vma is no longer available, don't continue to swapin */
 924                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 925                                 return false;
 926                         }
 927                         /* check if the pmd is still valid */
 928                         if (mm_find_pmd(mm, address) != pmd) {
 929                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 930                                 return false;
 931                         }
 932                 }
 933                 if (ret & VM_FAULT_ERROR) {
 934                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 935                         return false;
 936                 }
 937                 /* pte is unmapped now, we need to map it */
 938                 vmf.pte = pte_offset_map(pmd, vmf.address);
 939         }
 940         vmf.pte--;
 941         pte_unmap(vmf.pte);
 942         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
 943         return true;
 944 }
 945
 946 static void collapse_huge_page(struct mm_struct *mm,
 947                                    unsigned long address,
 948                                    struct page **hpage,
 949                                    int node, int referenced)
 950 {
 951         pmd_t *pmd, _pmd;
 952         pte_t *pte;
 953         pgtable_t pgtable;
 954         struct page *new_page;
 955         spinlock_t *pmd_ptl, *pte_ptl;
 956         int isolated = 0, result = 0;
 957         struct mem_cgroup *memcg;
 958         struct vm_area_struct *vma;
 959         struct mmu_notifier_range range;
 960         gfp_t gfp;
 961
 962         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 963
 964         /* Only allocate from the target node */
 965         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
 966
 967         /*
 968          * Before allocating the hugepage, release the mmap_sem read lock.
 969          * The allocation can take potentially a long time if it involves
 970          * sync compaction, and we do not need to hold the mmap_sem during
 971          * that. We will recheck the vma after taking it again in write mode.
 972          */
 973         up_read(&mm->mmap_sem);
 974         new_page = khugepaged_alloc_page(hpage, gfp, node);
 975         if (!new_page) {
 976                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
 977                 goto out_nolock;
 978         }
 979
 980         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
 981                 result = SCAN_CGROUP_CHARGE_FAIL;
 982                 goto out_nolock;
 983         }
 984
 985         down_read(&mm->mmap_sem);
 986         result = hugepage_vma_revalidate(mm, address, &vma);
 987         if (result) {
 988                 mem_cgroup_cancel_charge(new_page, memcg, true);
 989                 up_read(&mm->mmap_sem);
 990                 goto out_nolock;
 991         }
 992
 993         pmd = mm_find_pmd(mm, address);
 994         if (!pmd) {
 995                 result = SCAN_PMD_NULL;
 996                 mem_cgroup_cancel_charge(new_page, memcg, true);
 997                 up_read(&mm->mmap_sem);
 998                 goto out_nolock;
 999         }
1000
1001         /*
1002          * __collapse_huge_page_swapin always returns with mmap_sem locked.
1003          * If it fails, we release mmap_sem and jump out_nolock.
1004          * Continuing to collapse causes inconsistency.
1005          */
1006         if (!__collapse_huge_page_swapin(mm, vma, address, pmd, referenced)) {
1007                 mem_cgroup_cancel_charge(new_page, memcg, true);
1008                 up_read(&mm->mmap_sem);
1009                 goto out_nolock;
1010         }
1011
1012         up_read(&mm->mmap_sem);
1013         /*
1014          * Prevent all access to pagetables with the exception of
1015          * gup_fast later handled by the ptep_clear_flush and the VM
1016          * handled by the anon_vma lock + PG_lock.
1017          */
1018         down_write(&mm->mmap_sem);
1019         result = SCAN_ANY_PROCESS;
1020         if (!mmget_still_valid(mm))
1021                 goto out;
1022         result = hugepage_vma_revalidate(mm, address, &vma);
1023         if (result)
1024                 goto out;
1025         /* check if the pmd is still valid */
1026         if (mm_find_pmd(mm, address) != pmd)
1027                 goto out;
1028
1029         anon_vma_lock_write(vma->anon_vma);
1030
1031         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1032                                 address, address + HPAGE_PMD_SIZE);
1033         mmu_notifier_invalidate_range_start(&range);
1034
1035         pte = pte_offset_map(pmd, address);
1036         pte_ptl = pte_lockptr(mm, pmd);
1037
1038         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1039         /*
1040          * After this gup_fast can't run anymore. This also removes
1041          * any huge TLB entry from the CPU so we won't allow
1042          * huge and small TLB entries for the same virtual address
1043          * to avoid the risk of CPU bugs in that area.
1044          */
1045         _pmd = pmdp_collapse_flush(vma, address, pmd);
1046         spin_unlock(pmd_ptl);
1047         mmu_notifier_invalidate_range_end(&range);
1048
1049         spin_lock(pte_ptl);
1050         isolated = __collapse_huge_page_isolate(vma, address, pte);
1051         spin_unlock(pte_ptl);
1052
1053         if (unlikely(!isolated)) {
1054                 pte_unmap(pte);
1055                 spin_lock(pmd_ptl);
1056                 BUG_ON(!pmd_none(*pmd));
1057                 /*
1058                  * We can only use set_pmd_at when establishing
1059                  * hugepmds and never for establishing regular pmds that
1060                  * points to regular pagetables. Use pmd_populate for that
1061                  */
1062                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1063                 spin_unlock(pmd_ptl);
1064                 anon_vma_unlock_write(vma->anon_vma);
1065                 result = SCAN_FAIL;
1066                 goto out;
1067         }
1068
1069         /*
1070          * All pages are isolated and locked so anon_vma rmap
1071          * can't run anymore.
1072          */
1073         anon_vma_unlock_write(vma->anon_vma);
1074
1075         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
1076         pte_unmap(pte);
1077         __SetPageUptodate(new_page);
1078         pgtable = pmd_pgtable(_pmd);
1079
1080         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1081         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1082
1083         /*
1084          * spin_lock() below is not the equivalent of smp_wmb(), so
1085          * this is needed to avoid the copy_huge_page writes to become
1086          * visible after the set_pmd_at() write.
1087          */
1088         smp_wmb();
1089
1090         spin_lock(pmd_ptl);
1091         BUG_ON(!pmd_none(*pmd));
1092         page_add_new_anon_rmap(new_page, vma, address, true);
1093         mem_cgroup_commit_charge(new_page, memcg, false, true);
1094         count_memcg_events(memcg, THP_COLLAPSE_ALLOC, 1);
1095         lru_cache_add_active_or_unevictable(new_page, vma);
1096         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1097         set_pmd_at(mm, address, pmd, _pmd);
1098         update_mmu_cache_pmd(vma, address, pmd);
1099         spin_unlock(pmd_ptl);
1100
1101         *hpage = NULL;
1102
1103         khugepaged_pages_collapsed++;
1104         result = SCAN_SUCCEED;
1105 out_up_write:
1106         up_write(&mm->mmap_sem);
1107 out_nolock:
1108         trace_mm_collapse_huge_page(mm, isolated, result);
1109         return;
1110 out:
1111         mem_cgroup_cancel_charge(new_page, memcg, true);
1112         goto out_up_write;
1113 }
1114
1115 static int khugepaged_scan_pmd(struct mm_struct *mm,
1116                                struct vm_area_struct *vma,
1117                                unsigned long address,
1118                                struct page **hpage)
1119 {
1120         pmd_t *pmd;
1121         pte_t *pte, *_pte;
1122         int ret = 0, none_or_zero = 0, result = 0, referenced = 0;
1123         struct page *page = NULL;
1124         unsigned long _address;
1125         spinlock_t *ptl;
1126         int node = NUMA_NO_NODE, unmapped = 0;
1127         bool writable = false;
1128
1129         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1130
1131         pmd = mm_find_pmd(mm, address);
1132         if (!pmd) {
1133                 result = SCAN_PMD_NULL;
1134                 goto out;
1135         }
1136
1137         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1138         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1139         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1140              _pte++, _address += PAGE_SIZE) {
1141                 pte_t pteval = *_pte;
1142                 if (is_swap_pte(pteval)) {
1143                         if (++unmapped <= khugepaged_max_ptes_swap) {
1144                                 continue;
1145                         } else {
1146                                 result = SCAN_EXCEED_SWAP_PTE;
1147                                 goto out_unmap;
1148                         }
1149                 }
1150                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1151                         if (!userfaultfd_armed(vma) &&
1152                             ++none_or_zero <= khugepaged_max_ptes_none) {
1153                                 continue;
1154                         } else {
1155                                 result = SCAN_EXCEED_NONE_PTE;
1156                                 goto out_unmap;
1157                         }
1158                 }
1159                 if (!pte_present(pteval)) {
1160                         result = SCAN_PTE_NON_PRESENT;
1161                         goto out_unmap;
1162                 }
1163                 if (pte_write(pteval))
1164                         writable = true;
1165
1166                 page = vm_normal_page(vma, _address, pteval);
1167                 if (unlikely(!page)) {
1168                         result = SCAN_PAGE_NULL;
1169                         goto out_unmap;
1170                 }
1171
1172                 /* TODO: teach khugepaged to collapse THP mapped with pte */
1173                 if (PageCompound(page)) {
1174                         result = SCAN_PAGE_COMPOUND;
1175                         goto out_unmap;
1176                 }
1177
1178                 /*
1179                  * Record which node the original page is from and save this
1180                  * information to khugepaged_node_load[].
1181                  * Khupaged will allocate hugepage from the node has the max
1182                  * hit record.
1183                  */
1184                 node = page_to_nid(page);
1185                 if (khugepaged_scan_abort(node)) {
1186                         result = SCAN_SCAN_ABORT;
1187                         goto out_unmap;
1188                 }
1189                 khugepaged_node_load[node]++;
1190                 if (!PageLRU(page)) {
1191                         result = SCAN_PAGE_LRU;
1192                         goto out_unmap;
1193                 }
1194                 if (PageLocked(page)) {
1195                         result = SCAN_PAGE_LOCK;
1196                         goto out_unmap;
1197                 }
1198                 if (!PageAnon(page)) {
1199                         result = SCAN_PAGE_ANON;
1200                         goto out_unmap;
1201                 }
1202
1203                 /*
1204                  * cannot use mapcount: can't collapse if there's a gup pin.
1205                  * The page must only be referenced by the scanned process
1206                  * and page swap cache.
1207                  */
1208                 if (page_count(page) != 1 + PageSwapCache(page)) {
1209                         result = SCAN_PAGE_COUNT;
1210                         goto out_unmap;
1211                 }
1212                 if (pte_young(pteval) ||
1213                     page_is_young(page) || PageReferenced(page) ||
1214                     mmu_notifier_test_young(vma->vm_mm, address))
1215                         referenced++;
1216         }
1217         if (writable) {
1218                 if (referenced) {
1219                         result = SCAN_SUCCEED;
1220                         ret = 1;
1221                 } else {
1222                         result = SCAN_LACK_REFERENCED_PAGE;
1223                 }
1224         } else {
1225                 result = SCAN_PAGE_RO;
1226         }
1227 out_unmap:
1228         pte_unmap_unlock(pte, ptl);
1229         if (ret) {
1230                 node = khugepaged_find_target_node();
1231                 /* collapse_huge_page will return with the mmap_sem released */
1232                 collapse_huge_page(mm, address, hpage, node, referenced);
1233         }
1234 out:
1235         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1236                                      none_or_zero, result, unmapped);
1237         return ret;
1238 }
1239
1240 static void collect_mm_slot(struct mm_slot *mm_slot)
1241 {
1242         struct mm_struct *mm = mm_slot->mm;
1243
1244         lockdep_assert_held(&khugepaged_mm_lock);
1245
1246         if (khugepaged_test_exit(mm)) {
1247                 /* free mm_slot */
1248                 hash_del(&mm_slot->hash);
1249                 list_del(&mm_slot->mm_node);
1250
1251                 /*
1252                  * Not strictly needed because the mm exited already.
1253                  *
1254                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1255                  */
1256
1257                 /* khugepaged_mm_lock actually not necessary for the below */
1258                 free_mm_slot(mm_slot);
1259                 mmdrop(mm);
1260         }
1261 }
1262
1263 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
1264 /*
1265  * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1266  * khugepaged should try to collapse the page table.
1267  */
1268 static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1269                                          unsigned long addr)
1270 {
1271         struct mm_slot *mm_slot;
1272
1273         VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1274
1275         spin_lock(&khugepaged_mm_lock);
1276         mm_slot = get_mm_slot(mm);
1277         if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1278                 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1279         spin_unlock(&khugepaged_mm_lock);
1280         return 0;
1281 }
1282
1283 /**
1284  * Try to collapse a pte-mapped THP for mm at address haddr.
1285  *
1286  * This function checks whether all the PTEs in the PMD are pointing to the
1287  * right THP. If so, retract the page table so the THP can refault in with
1288  * as pmd-mapped.
1289  */
1290 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1291 {
1292         unsigned long haddr = addr & HPAGE_PMD_MASK;
1293         struct vm_area_struct *vma = find_vma(mm, haddr);
1294         struct page *hpage = NULL;
1295         pte_t *start_pte, *pte;
1296         pmd_t *pmd, _pmd;
1297         spinlock_t *ptl;
1298         int count = 0;
1299         int i;
1300
1301         if (!vma || !vma->vm_file ||
1302             vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE)
1303                 return;
1304
1305         /*
1306          * This vm_flags may not have VM_HUGEPAGE if the page was not
1307          * collapsed by this mm. But we can still collapse if the page is
1308          * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1309          * will not fail the vma for missing VM_HUGEPAGE
1310          */
1311         if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE))
1312                 return;
1313
1314         pmd = mm_find_pmd(mm, haddr);
1315         if (!pmd)
1316                 return;
1317
1318         start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1319
1320         /* step 1: check all mapped PTEs are to the right huge page */
1321         for (i = 0, addr = haddr, pte = start_pte;
1322              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1323                 struct page *page;
1324
1325                 /* empty pte, skip */
1326                 if (pte_none(*pte))
1327                         continue;
1328
1329                 /* page swapped out, abort */
1330                 if (!pte_present(*pte))
1331                         goto abort;
1332
1333                 page = vm_normal_page(vma, addr, *pte);
1334
1335                 if (!page || !PageCompound(page))
1336                         goto abort;
1337
1338                 if (!hpage) {
1339                         hpage = compound_head(page);
1340                         /*
1341                          * The mapping of the THP should not change.
1342                          *
1343                          * Note that uprobe, debugger, or MAP_PRIVATE may
1344                          * change the page table, but the new page will
1345                          * not pass PageCompound() check.
1346                          */
1347                         if (WARN_ON(hpage->mapping != vma->vm_file->f_mapping))
1348                                 goto abort;
1349                 }
1350
1351                 /*
1352                  * Confirm the page maps to the correct subpage.
1353                  *
1354                  * Note that uprobe, debugger, or MAP_PRIVATE may change
1355                  * the page table, but the new page will not pass
1356                  * PageCompound() check.
1357                  */
1358                 if (WARN_ON(hpage + i != page))
1359                         goto abort;
1360                 count++;
1361         }
1362
1363         /* step 2: adjust rmap */
1364         for (i = 0, addr = haddr, pte = start_pte;
1365              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1366                 struct page *page;
1367
1368                 if (pte_none(*pte))
1369                         continue;
1370                 page = vm_normal_page(vma, addr, *pte);
1371                 page_remove_rmap(page, false);
1372         }
1373
1374         pte_unmap_unlock(start_pte, ptl);
1375
1376         /* step 3: set proper refcount and mm_counters. */
1377         if (hpage) {
1378                 page_ref_sub(hpage, count);
1379                 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1380         }
1381
1382         /* step 4: collapse pmd */
1383         ptl = pmd_lock(vma->vm_mm, pmd);
1384         _pmd = pmdp_collapse_flush(vma, addr, pmd);
1385         spin_unlock(ptl);
1386         mm_dec_nr_ptes(mm);
1387         pte_free(mm, pmd_pgtable(_pmd));
1388         return;
1389
1390 abort:
1391         pte_unmap_unlock(start_pte, ptl);
1392 }
1393
1394 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1395 {
1396         struct mm_struct *mm = mm_slot->mm;
1397         int i;
1398
1399         if (likely(mm_slot->nr_pte_mapped_thp == 0))
1400                 return 0;
1401
1402         if (!down_write_trylock(&mm->mmap_sem))
1403                 return -EBUSY;
1404
1405         if (unlikely(khugepaged_test_exit(mm)))
1406                 goto out;
1407
1408         for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1409                 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1410
1411 out:
1412         mm_slot->nr_pte_mapped_thp = 0;
1413         up_write(&mm->mmap_sem);
1414         return 0;
1415 }
1416
1417 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1418 {
1419         struct vm_area_struct *vma;
1420         unsigned long addr;
1421         pmd_t *pmd, _pmd;
1422
1423         i_mmap_lock_write(mapping);
1424         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1425                 /*
1426                  * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1427                  * got written to. These VMAs are likely not worth investing
1428                  * down_write(mmap_sem) as PMD-mapping is likely to be split
1429                  * later.
1430                  *
1431                  * Not that vma->anon_vma check is racy: it can be set up after
1432                  * the check but before we took mmap_sem by the fault path.
1433                  * But page lock would prevent establishing any new ptes of the
1434                  * page, so we are safe.
1435                  *
1436                  * An alternative would be drop the check, but check that page
1437                  * table is clear before calling pmdp_collapse_flush() under
1438                  * ptl. It has higher chance to recover THP for the VMA, but
1439                  * has higher cost too.
1440                  */
1441                 if (vma->anon_vma)
1442                         continue;
1443                 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1444                 if (addr & ~HPAGE_PMD_MASK)
1445                         continue;
1446                 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1447                         continue;
1448                 pmd = mm_find_pmd(vma->vm_mm, addr);
1449                 if (!pmd)
1450                         continue;
1451                 /*
1452                  * We need exclusive mmap_sem to retract page table.
1453                  *
1454                  * We use trylock due to lock inversion: we need to acquire
1455                  * mmap_sem while holding page lock. Fault path does it in
1456                  * reverse order. Trylock is a way to avoid deadlock.
1457                  */
1458                 if (down_write_trylock(&vma->vm_mm->mmap_sem)) {
1459                         spinlock_t *ptl = pmd_lock(vma->vm_mm, pmd);
1460                         /* assume page table is clear */
1461                         _pmd = pmdp_collapse_flush(vma, addr, pmd);
1462                         spin_unlock(ptl);
1463                         up_write(&vma->vm_mm->mmap_sem);
1464                         mm_dec_nr_ptes(vma->vm_mm);
1465                         pte_free(vma->vm_mm, pmd_pgtable(_pmd));
1466                 } else {
1467                         /* Try again later */
1468                         khugepaged_add_pte_mapped_thp(vma->vm_mm, addr);
1469                 }
1470         }
1471         i_mmap_unlock_write(mapping);
1472 }
1473
1474 /**
1475  * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1476  *
1477  * Basic scheme is simple, details are more complex:
1478  *  - allocate and lock a new huge page;
1479  *  - scan page cache replacing old pages with the new one
1480  *    + swap/gup in pages if necessary;
1481  *    + fill in gaps;
1482  *    + keep old pages around in case rollback is required;
1483  *  - if replacing succeeds:
1484  *    + copy data over;
1485  *    + free old pages;
1486  *    + unlock huge page;
1487  *  - if replacing failed;
1488  *    + put all pages back and unfreeze them;
1489  *    + restore gaps in the page cache;
1490  *    + unlock and free huge page;
1491  */
1492 static void collapse_file(struct mm_struct *mm,
1493                 struct file *file, pgoff_t start,
1494                 struct page **hpage, int node)
1495 {
1496         struct address_space *mapping = file->f_mapping;
1497         gfp_t gfp;
1498         struct page *new_page;
1499         struct mem_cgroup *memcg;
1500         pgoff_t index, end = start + HPAGE_PMD_NR;
1501         LIST_HEAD(pagelist);
1502         XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1503         int nr_none = 0, result = SCAN_SUCCEED;
1504         bool is_shmem = shmem_file(file);
1505
1506         VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1507         VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1508
1509         /* Only allocate from the target node */
1510         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1511
1512         new_page = khugepaged_alloc_page(hpage, gfp, node);
1513         if (!new_page) {
1514                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1515                 goto out;
1516         }
1517
1518         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
1519                 result = SCAN_CGROUP_CHARGE_FAIL;
1520                 goto out;
1521         }
1522
1523         /* This will be less messy when we use multi-index entries */
1524         do {
1525                 xas_lock_irq(&xas);
1526                 xas_create_range(&xas);
1527                 if (!xas_error(&xas))
1528                         break;
1529                 xas_unlock_irq(&xas);
1530                 if (!xas_nomem(&xas, GFP_KERNEL)) {
1531                         mem_cgroup_cancel_charge(new_page, memcg, true);
1532                         result = SCAN_FAIL;
1533                         goto out;
1534                 }
1535         } while (1);
1536
1537         __SetPageLocked(new_page);
1538         if (is_shmem)
1539                 __SetPageSwapBacked(new_page);
1540         new_page->index = start;
1541         new_page->mapping = mapping;
1542
1543         /*
1544          * At this point the new_page is locked and not up-to-date.
1545          * It's safe to insert it into the page cache, because nobody would
1546          * be able to map it or use it in another way until we unlock it.
1547          */
1548
1549         xas_set(&xas, start);
1550         for (index = start; index < end; index++) {
1551                 struct page *page = xas_next(&xas);
1552
1553                 VM_BUG_ON(index != xas.xa_index);
1554                 if (is_shmem) {
1555                         if (!page) {
1556                                 /*
1557                                  * Stop if extent has been truncated or
1558                                  * hole-punched, and is now completely
1559                                  * empty.
1560                                  */
1561                                 if (index == start) {
1562                                         if (!xas_next_entry(&xas, end - 1)) {
1563                                                 result = SCAN_TRUNCATED;
1564                                                 goto xa_locked;
1565                                         }
1566                                         xas_set(&xas, index);
1567                                 }
1568                                 if (!shmem_charge(mapping->host, 1)) {
1569                                         result = SCAN_FAIL;
1570                                         goto xa_locked;
1571                                 }
1572                                 xas_store(&xas, new_page);
1573                                 nr_none++;
1574                                 continue;
1575                         }
1576
1577                         if (xa_is_value(page) || !PageUptodate(page)) {
1578                                 xas_unlock_irq(&xas);
1579                                 /* swap in or instantiate fallocated page */
1580                                 if (shmem_getpage(mapping->host, index, &page,
1581                                                   SGP_NOHUGE)) {
1582                                         result = SCAN_FAIL;
1583                                         goto xa_unlocked;
1584                                 }
1585                         } else if (trylock_page(page)) {
1586                                 get_page(page);
1587                                 xas_unlock_irq(&xas);
1588                         } else {
1589                                 result = SCAN_PAGE_LOCK;
1590                                 goto xa_locked;
1591                         }
1592                 } else {        /* !is_shmem */
1593                         if (!page || xa_is_value(page)) {
1594                                 xas_unlock_irq(&xas);
1595                                 page_cache_sync_readahead(mapping, &file->f_ra,
1596                                                           file, index,
1597                                                           PAGE_SIZE);
1598                                 /* drain pagevecs to help isolate_lru_page() */
1599                                 lru_add_drain();
1600                                 page = find_lock_page(mapping, index);
1601                                 if (unlikely(page == NULL)) {
1602                                         result = SCAN_FAIL;
1603                                         goto xa_unlocked;
1604                                 }
1605                         } else if (PageDirty(page)) {
1606                                 /*
1607                                  * khugepaged only works on read-only fd,
1608                                  * so this page is dirty because it hasn't
1609                                  * been flushed since first write. There
1610                                  * won't be new dirty pages.
1611                                  *
1612                                  * Trigger async flush here and hope the
1613                                  * writeback is done when khugepaged
1614                                  * revisits this page.
1615                                  *
1616                                  * This is a one-off situation. We are not
1617                                  * forcing writeback in loop.
1618                                  */
1619                                 xas_unlock_irq(&xas);
1620                                 filemap_flush(mapping);
1621                                 result = SCAN_FAIL;
1622                                 goto xa_unlocked;
1623                         } else if (trylock_page(page)) {
1624                                 get_page(page);
1625                                 xas_unlock_irq(&xas);
1626                         } else {
1627                                 result = SCAN_PAGE_LOCK;
1628                                 goto xa_locked;
1629                         }
1630                 }
1631
1632                 /*
1633                  * The page must be locked, so we can drop the i_pages lock
1634                  * without racing with truncate.
1635                  */
1636                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1637
1638                 /* make sure the page is up to date */
1639                 if (unlikely(!PageUptodate(page))) {
1640                         result = SCAN_FAIL;
1641                         goto out_unlock;
1642                 }
1643
1644                 /*
1645                  * If file was truncated then extended, or hole-punched, before
1646                  * we locked the first page, then a THP might be there already.
1647                  */
1648                 if (PageTransCompound(page)) {
1649                         result = SCAN_PAGE_COMPOUND;
1650                         goto out_unlock;
1651                 }
1652
1653                 if (page_mapping(page) != mapping) {
1654                         result = SCAN_TRUNCATED;
1655                         goto out_unlock;
1656                 }
1657
1658                 if (!is_shmem && PageDirty(page)) {
1659                         /*
1660                          * khugepaged only works on read-only fd, so this
1661                          * page is dirty because it hasn't been flushed
1662                          * since first write.
1663                          */
1664                         result = SCAN_FAIL;
1665                         goto out_unlock;
1666                 }
1667
1668                 if (isolate_lru_page(page)) {
1669                         result = SCAN_DEL_PAGE_LRU;
1670                         goto out_unlock;
1671                 }
1672
1673                 if (page_has_private(page) &&
1674                     !try_to_release_page(page, GFP_KERNEL)) {
1675                         result = SCAN_PAGE_HAS_PRIVATE;
1676                         goto out_unlock;
1677                 }
1678
1679                 if (page_mapped(page))
1680                         unmap_mapping_pages(mapping, index, 1, false);
1681
1682                 xas_lock_irq(&xas);
1683                 xas_set(&xas, index);
1684
1685                 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1686                 VM_BUG_ON_PAGE(page_mapped(page), page);
1687
1688                 /*
1689                  * The page is expected to have page_count() == 3:
1690                  *  - we hold a pin on it;
1691                  *  - one reference from page cache;
1692                  *  - one from isolate_lru_page;
1693                  */
1694                 if (!page_ref_freeze(page, 3)) {
1695                         result = SCAN_PAGE_COUNT;
1696                         xas_unlock_irq(&xas);
1697                         putback_lru_page(page);
1698                         goto out_unlock;
1699                 }
1700
1701                 /*
1702                  * Add the page to the list to be able to undo the collapse if
1703                  * something go wrong.
1704                  */
1705                 list_add_tail(&page->lru, &pagelist);
1706
1707                 /* Finally, replace with the new page. */
1708                 xas_store(&xas, new_page);
1709                 continue;
1710 out_unlock:
1711                 unlock_page(page);
1712                 put_page(page);
1713                 goto xa_unlocked;
1714         }
1715
1716         if (is_shmem)
1717                 __inc_node_page_state(new_page, NR_SHMEM_THPS);
1718         else {
1719                 __inc_node_page_state(new_page, NR_FILE_THPS);
1720                 filemap_nr_thps_inc(mapping);
1721         }
1722
1723         if (nr_none) {
1724                 struct zone *zone = page_zone(new_page);
1725
1726                 __mod_node_page_state(zone->zone_pgdat, NR_FILE_PAGES, nr_none);
1727                 if (is_shmem)
1728                         __mod_node_page_state(zone->zone_pgdat,
1729                                               NR_SHMEM, nr_none);
1730         }
1731
1732 xa_locked:
1733         xas_unlock_irq(&xas);
1734 xa_unlocked:
1735
1736         if (result == SCAN_SUCCEED) {
1737                 struct page *page, *tmp;
1738
1739                 /*
1740                  * Replacing old pages with new one has succeeded, now we
1741                  * need to copy the content and free the old pages.
1742                  */
1743                 index = start;
1744                 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1745                         while (index < page->index) {
1746                                 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1747                                 index++;
1748                         }
1749                         copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1750                                         page);
1751                         list_del(&page->lru);
1752                         page->mapping = NULL;
1753                         page_ref_unfreeze(page, 1);
1754                         ClearPageActive(page);
1755                         ClearPageUnevictable(page);
1756                         unlock_page(page);
1757                         put_page(page);
1758                         index++;
1759                 }
1760                 while (index < end) {
1761                         clear_highpage(new_page + (index % HPAGE_PMD_NR));
1762                         index++;
1763                 }
1764
1765                 SetPageUptodate(new_page);
1766                 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1767                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1768
1769                 if (is_shmem) {
1770                         set_page_dirty(new_page);
1771                         lru_cache_add_anon(new_page);
1772                 } else {
1773                         lru_cache_add_file(new_page);
1774                 }
1775                 count_memcg_events(memcg, THP_COLLAPSE_ALLOC, 1);
1776
1777                 /*
1778                  * Remove pte page tables, so we can re-fault the page as huge.
1779                  */
1780                 retract_page_tables(mapping, start);
1781                 *hpage = NULL;
1782
1783                 khugepaged_pages_collapsed++;
1784         } else {
1785                 struct page *page;
1786
1787                 /* Something went wrong: roll back page cache changes */
1788                 xas_lock_irq(&xas);
1789                 mapping->nrpages -= nr_none;
1790
1791                 if (is_shmem)
1792                         shmem_uncharge(mapping->host, nr_none);
1793
1794                 xas_set(&xas, start);
1795                 xas_for_each(&xas, page, end - 1) {
1796                         page = list_first_entry_or_null(&pagelist,
1797                                         struct page, lru);
1798                         if (!page || xas.xa_index < page->index) {
1799                                 if (!nr_none)
1800                                         break;
1801                                 nr_none--;
1802                                 /* Put holes back where they were */
1803                                 xas_store(&xas, NULL);
1804                                 continue;
1805                         }
1806
1807                         VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1808
1809                         /* Unfreeze the page. */
1810                         list_del(&page->lru);
1811                         page_ref_unfreeze(page, 2);
1812                         xas_store(&xas, page);
1813                         xas_pause(&xas);
1814                         xas_unlock_irq(&xas);
1815                         unlock_page(page);
1816                         putback_lru_page(page);
1817                         xas_lock_irq(&xas);
1818                 }
1819                 VM_BUG_ON(nr_none);
1820                 xas_unlock_irq(&xas);
1821
1822                 mem_cgroup_cancel_charge(new_page, memcg, true);
1823                 new_page->mapping = NULL;
1824         }
1825
1826         unlock_page(new_page);
1827 out:
1828         VM_BUG_ON(!list_empty(&pagelist));
1829         /* TODO: tracepoints */
1830 }
1831
1832 static void khugepaged_scan_file(struct mm_struct *mm,
1833                 struct file *file, pgoff_t start, struct page **hpage)
1834 {
1835         struct page *page = NULL;
1836         struct address_space *mapping = file->f_mapping;
1837         XA_STATE(xas, &mapping->i_pages, start);
1838         int present, swap;
1839         int node = NUMA_NO_NODE;
1840         int result = SCAN_SUCCEED;
1841
1842         present = 0;
1843         swap = 0;
1844         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1845         rcu_read_lock();
1846         xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
1847                 if (xas_retry(&xas, page))
1848                         continue;
1849
1850                 if (xa_is_value(page)) {
1851                         if (++swap > khugepaged_max_ptes_swap) {
1852                                 result = SCAN_EXCEED_SWAP_PTE;
1853                                 break;
1854                         }
1855                         continue;
1856                 }
1857
1858                 if (PageTransCompound(page)) {
1859                         result = SCAN_PAGE_COMPOUND;
1860                         break;
1861                 }
1862
1863                 node = page_to_nid(page);
1864                 if (khugepaged_scan_abort(node)) {
1865                         result = SCAN_SCAN_ABORT;
1866                         break;
1867                 }
1868                 khugepaged_node_load[node]++;
1869
1870                 if (!PageLRU(page)) {
1871                         result = SCAN_PAGE_LRU;
1872                         break;
1873                 }
1874
1875                 if (page_count(page) !=
1876                     1 + page_mapcount(page) + page_has_private(page)) {
1877                         result = SCAN_PAGE_COUNT;
1878                         break;
1879                 }
1880
1881                 /*
1882                  * We probably should check if the page is referenced here, but
1883                  * nobody would transfer pte_young() to PageReferenced() for us.
1884                  * And rmap walk here is just too costly...
1885                  */
1886
1887                 present++;
1888
1889                 if (need_resched()) {
1890                         xas_pause(&xas);
1891                         cond_resched_rcu();
1892                 }
1893         }
1894         rcu_read_unlock();
1895
1896         if (result == SCAN_SUCCEED) {
1897                 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
1898                         result = SCAN_EXCEED_NONE_PTE;
1899                 } else {
1900                         node = khugepaged_find_target_node();
1901                         collapse_file(mm, file, start, hpage, node);
1902                 }
1903         }
1904
1905         /* TODO: tracepoints */
1906 }
1907 #else
1908 static void khugepaged_scan_file(struct mm_struct *mm,
1909                 struct file *file, pgoff_t start, struct page **hpage)
1910 {
1911         BUILD_BUG();
1912 }
1913
1914 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1915 {
1916         return 0;
1917 }
1918 #endif
1919
1920 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
1921                                             struct page **hpage)
1922         __releases(&khugepaged_mm_lock)
1923         __acquires(&khugepaged_mm_lock)
1924 {
1925         struct mm_slot *mm_slot;
1926         struct mm_struct *mm;
1927         struct vm_area_struct *vma;
1928         int progress = 0;
1929
1930         VM_BUG_ON(!pages);
1931         lockdep_assert_held(&khugepaged_mm_lock);
1932
1933         if (khugepaged_scan.mm_slot)
1934                 mm_slot = khugepaged_scan.mm_slot;
1935         else {
1936                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
1937                                      struct mm_slot, mm_node);
1938                 khugepaged_scan.address = 0;
1939                 khugepaged_scan.mm_slot = mm_slot;
1940         }
1941         spin_unlock(&khugepaged_mm_lock);
1942         khugepaged_collapse_pte_mapped_thps(mm_slot);
1943
1944         mm = mm_slot->mm;
1945         /*
1946          * Don't wait for semaphore (to avoid long wait times).  Just move to
1947          * the next mm on the list.
1948          */
1949         vma = NULL;
1950         if (unlikely(!down_read_trylock(&mm->mmap_sem)))
1951                 goto breakouterloop_mmap_sem;
1952         if (likely(!khugepaged_test_exit(mm)))
1953                 vma = find_vma(mm, khugepaged_scan.address);
1954
1955         progress++;
1956         for (; vma; vma = vma->vm_next) {
1957                 unsigned long hstart, hend;
1958
1959                 cond_resched();
1960                 if (unlikely(khugepaged_test_exit(mm))) {
1961                         progress++;
1962                         break;
1963                 }
1964                 if (!hugepage_vma_check(vma, vma->vm_flags)) {
1965 skip:
1966                         progress++;
1967                         continue;
1968                 }
1969                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1970                 hend = vma->vm_end & HPAGE_PMD_MASK;
1971                 if (hstart >= hend)
1972                         goto skip;
1973                 if (khugepaged_scan.address > hend)
1974                         goto skip;
1975                 if (khugepaged_scan.address < hstart)
1976                         khugepaged_scan.address = hstart;
1977                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
1978
1979                 while (khugepaged_scan.address < hend) {
1980                         int ret;
1981                         cond_resched();
1982                         if (unlikely(khugepaged_test_exit(mm)))
1983                                 goto breakouterloop;
1984
1985                         VM_BUG_ON(khugepaged_scan.address < hstart ||
1986                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
1987                                   hend);
1988                         if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
1989                                 struct file *file;
1990                                 pgoff_t pgoff = linear_page_index(vma,
1991                                                 khugepaged_scan.address);
1992
1993                                 if (shmem_file(vma->vm_file)
1994                                     && !shmem_huge_enabled(vma))
1995                                         goto skip;
1996                                 file = get_file(vma->vm_file);
1997                                 up_read(&mm->mmap_sem);
1998                                 ret = 1;
1999                                 khugepaged_scan_file(mm, file, pgoff, hpage);
2000                                 fput(file);
2001                         } else {
2002                                 ret = khugepaged_scan_pmd(mm, vma,
2003                                                 khugepaged_scan.address,
2004                                                 hpage);
2005                         }
2006                         /* move to next address */
2007                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2008                         progress += HPAGE_PMD_NR;
2009                         if (ret)
2010                                 /* we released mmap_sem so break loop */
2011                                 goto breakouterloop_mmap_sem;
2012                         if (progress >= pages)
2013                                 goto breakouterloop;
2014                 }
2015         }
2016 breakouterloop:
2017         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2018 breakouterloop_mmap_sem:
2019
2020         spin_lock(&khugepaged_mm_lock);
2021         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2022         /*
2023          * Release the current mm_slot if this mm is about to die, or
2024          * if we scanned all vmas of this mm.
2025          */
2026         if (khugepaged_test_exit(mm) || !vma) {
2027                 /*
2028                  * Make sure that if mm_users is reaching zero while
2029                  * khugepaged runs here, khugepaged_exit will find
2030                  * mm_slot not pointing to the exiting mm.
2031                  */
2032                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2033                         khugepaged_scan.mm_slot = list_entry(
2034                                 mm_slot->mm_node.next,
2035                                 struct mm_slot, mm_node);
2036                         khugepaged_scan.address = 0;
2037                 } else {
2038                         khugepaged_scan.mm_slot = NULL;
2039                         khugepaged_full_scans++;
2040                 }
2041
2042                 collect_mm_slot(mm_slot);
2043         }
2044
2045         return progress;
2046 }
2047
2048 static int khugepaged_has_work(void)
2049 {
2050         return !list_empty(&khugepaged_scan.mm_head) &&
2051                 khugepaged_enabled();
2052 }
2053
2054 static int khugepaged_wait_event(void)
2055 {
2056         return !list_empty(&khugepaged_scan.mm_head) ||
2057                 kthread_should_stop();
2058 }
2059
2060 static void khugepaged_do_scan(void)
2061 {
2062         struct page *hpage = NULL;
2063         unsigned int progress = 0, pass_through_head = 0;
2064         unsigned int pages = khugepaged_pages_to_scan;
2065         bool wait = true;
2066
2067         barrier(); /* write khugepaged_pages_to_scan to local stack */
2068
2069         while (progress < pages) {
2070                 if (!khugepaged_prealloc_page(&hpage, &wait))
2071                         break;
2072
2073                 cond_resched();
2074
2075                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2076                         break;
2077
2078                 spin_lock(&khugepaged_mm_lock);
2079                 if (!khugepaged_scan.mm_slot)
2080                         pass_through_head++;
2081                 if (khugepaged_has_work() &&
2082                     pass_through_head < 2)
2083                         progress += khugepaged_scan_mm_slot(pages - progress,
2084                                                             &hpage);
2085                 else
2086                         progress = pages;
2087                 spin_unlock(&khugepaged_mm_lock);
2088         }
2089
2090         if (!IS_ERR_OR_NULL(hpage))
2091                 put_page(hpage);
2092 }
2093
2094 static bool khugepaged_should_wakeup(void)
2095 {
2096         return kthread_should_stop() ||
2097                time_after_eq(jiffies, khugepaged_sleep_expire);
2098 }
2099
2100 static void khugepaged_wait_work(void)
2101 {
2102         if (khugepaged_has_work()) {
2103                 const unsigned long scan_sleep_jiffies =
2104                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2105
2106                 if (!scan_sleep_jiffies)
2107                         return;
2108
2109                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2110                 wait_event_freezable_timeout(khugepaged_wait,
2111                                              khugepaged_should_wakeup(),
2112                                              scan_sleep_jiffies);
2113                 return;
2114         }
2115
2116         if (khugepaged_enabled())
2117                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2118 }
2119
2120 static int khugepaged(void *none)
2121 {
2122         struct mm_slot *mm_slot;
2123
2124         set_freezable();
2125         set_user_nice(current, MAX_NICE);
2126
2127         while (!kthread_should_stop()) {
2128                 khugepaged_do_scan();
2129                 khugepaged_wait_work();
2130         }
2131
2132         spin_lock(&khugepaged_mm_lock);
2133         mm_slot = khugepaged_scan.mm_slot;
2134         khugepaged_scan.mm_slot = NULL;
2135         if (mm_slot)
2136                 collect_mm_slot(mm_slot);
2137         spin_unlock(&khugepaged_mm_lock);
2138         return 0;
2139 }
2140
2141 static void set_recommended_min_free_kbytes(void)
2142 {
2143         struct zone *zone;
2144         int nr_zones = 0;
2145         unsigned long recommended_min;
2146
2147         for_each_populated_zone(zone) {
2148                 /*
2149                  * We don't need to worry about fragmentation of
2150                  * ZONE_MOVABLE since it only has movable pages.
2151                  */
2152                 if (zone_idx(zone) > gfp_zone(GFP_USER))
2153                         continue;
2154
2155                 nr_zones++;
2156         }
2157
2158         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2159         recommended_min = pageblock_nr_pages * nr_zones * 2;
2160
2161         /*
2162          * Make sure that on average at least two pageblocks are almost free
2163          * of another type, one for a migratetype to fall back to and a
2164          * second to avoid subsequent fallbacks of other types There are 3
2165          * MIGRATE_TYPES we care about.
2166          */
2167         recommended_min += pageblock_nr_pages * nr_zones *
2168                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2169
2170         /* don't ever allow to reserve more than 5% of the lowmem */
2171         recommended_min = min(recommended_min,
2172                               (unsigned long) nr_free_buffer_pages() / 20);
2173         recommended_min <<= (PAGE_SHIFT-10);
2174
2175         if (recommended_min > min_free_kbytes) {
2176                 if (user_min_free_kbytes >= 0)
2177                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2178                                 min_free_kbytes, recommended_min);
2179
2180                 min_free_kbytes = recommended_min;
2181         }
2182         setup_per_zone_wmarks();
2183 }
2184
2185 int start_stop_khugepaged(void)
2186 {
2187         static struct task_struct *khugepaged_thread __read_mostly;
2188         static DEFINE_MUTEX(khugepaged_mutex);
2189         int err = 0;
2190
2191         mutex_lock(&khugepaged_mutex);
2192         if (khugepaged_enabled()) {
2193                 if (!khugepaged_thread)
2194                         khugepaged_thread = kthread_run(khugepaged, NULL,
2195                                                         "khugepaged");
2196                 if (IS_ERR(khugepaged_thread)) {
2197                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2198                         err = PTR_ERR(khugepaged_thread);
2199                         khugepaged_thread = NULL;
2200                         goto fail;
2201                 }
2202
2203                 if (!list_empty(&khugepaged_scan.mm_head))
2204                         wake_up_interruptible(&khugepaged_wait);
2205
2206                 set_recommended_min_free_kbytes();
2207         } else if (khugepaged_thread) {
2208                 kthread_stop(khugepaged_thread);
2209                 khugepaged_thread = NULL;
2210         }
2211 fail:
2212         mutex_unlock(&khugepaged_mutex);
2213         return err;
2214 }