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_PTE_UFFD_WP,
  33         SCAN_PAGE_RO,
  34         SCAN_LACK_REFERENCED_PAGE,
  35         SCAN_PAGE_NULL,
  36         SCAN_SCAN_ABORT,
  37         SCAN_PAGE_COUNT,
  38         SCAN_PAGE_LRU,
  39         SCAN_PAGE_LOCK,
  40         SCAN_PAGE_ANON,
  41         SCAN_PAGE_COMPOUND,
  42         SCAN_ANY_PROCESS,
  43         SCAN_VMA_NULL,
  44         SCAN_VMA_CHECK,
  45         SCAN_ADDRESS_RANGE,
  46         SCAN_SWAP_CACHE_PAGE,
  47         SCAN_DEL_PAGE_LRU,
  48         SCAN_ALLOC_HUGE_PAGE_FAIL,
  49         SCAN_CGROUP_CHARGE_FAIL,
  50         SCAN_EXCEED_SWAP_PTE,
  51         SCAN_TRUNCATED,
  52         SCAN_PAGE_HAS_PRIVATE,
  53 };
  54
  55 #define CREATE_TRACE_POINTS
  56 #include <trace/events/huge_memory.h>
  57
  58 /* default scan 8*512 pte (or vmas) every 30 second */
  59 static unsigned int khugepaged_pages_to_scan __read_mostly;
  60 static unsigned int khugepaged_pages_collapsed;
  61 static unsigned int khugepaged_full_scans;
  62 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  63 /* during fragmentation poll the hugepage allocator once every minute */
  64 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  65 static unsigned long khugepaged_sleep_expire;
  66 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  67 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  68 /*
  69  * default collapse hugepages if there is at least one pte mapped like
  70  * it would have happened if the vma was large enough during page
  71  * fault.
  72  */
  73 static unsigned int khugepaged_max_ptes_none __read_mostly;
  74 static unsigned int khugepaged_max_ptes_swap __read_mostly;
  75
  76 #define MM_SLOTS_HASH_BITS 10
  77 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  78
  79 static struct kmem_cache *mm_slot_cache __read_mostly;
  80
  81 #define MAX_PTE_MAPPED_THP 8
  82
  83 /**
  84  * struct mm_slot - hash lookup from mm to mm_slot
  85  * @hash: hash collision list
  86  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  87  * @mm: the mm that this information is valid for
  88  */
  89 struct mm_slot {
  90         struct hlist_node hash;
  91         struct list_head mm_node;
  92         struct mm_struct *mm;
  93
  94         /* pte-mapped THP in this mm */
  95         int nr_pte_mapped_thp;
  96         unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP];
  97 };
  98
  99 /**
 100  * struct khugepaged_scan - cursor for scanning
 101  * @mm_head: the head of the mm list to scan
 102  * @mm_slot: the current mm_slot we are scanning
 103  * @address: the next address inside that to be scanned
 104  *
 105  * There is only the one khugepaged_scan instance of this cursor structure.
 106  */
 107 struct khugepaged_scan {
 108         struct list_head mm_head;
 109         struct mm_slot *mm_slot;
 110         unsigned long address;
 111 };
 112
 113 static struct khugepaged_scan khugepaged_scan = {
 114         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 115 };
 116
 117 #ifdef CONFIG_SYSFS
 118 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 119                                          struct kobj_attribute *attr,
 120                                          char *buf)
 121 {
 122         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 123 }
 124
 125 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 126                                           struct kobj_attribute *attr,
 127                                           const char *buf, size_t count)
 128 {
 129         unsigned long msecs;
 130         int err;
 131
 132         err = kstrtoul(buf, 10, &msecs);
 133         if (err || msecs > UINT_MAX)
 134                 return -EINVAL;
 135
 136         khugepaged_scan_sleep_millisecs = msecs;
 137         khugepaged_sleep_expire = 0;
 138         wake_up_interruptible(&khugepaged_wait);
 139
 140         return count;
 141 }
 142 static struct kobj_attribute scan_sleep_millisecs_attr =
 143         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 144                scan_sleep_millisecs_store);
 145
 146 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 147                                           struct kobj_attribute *attr,
 148                                           char *buf)
 149 {
 150         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 151 }
 152
 153 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 154                                            struct kobj_attribute *attr,
 155                                            const char *buf, size_t count)
 156 {
 157         unsigned long msecs;
 158         int err;
 159
 160         err = kstrtoul(buf, 10, &msecs);
 161         if (err || msecs > UINT_MAX)
 162                 return -EINVAL;
 163
 164         khugepaged_alloc_sleep_millisecs = msecs;
 165         khugepaged_sleep_expire = 0;
 166         wake_up_interruptible(&khugepaged_wait);
 167
 168         return count;
 169 }
 170 static struct kobj_attribute alloc_sleep_millisecs_attr =
 171         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 172                alloc_sleep_millisecs_store);
 173
 174 static ssize_t pages_to_scan_show(struct kobject *kobj,
 175                                   struct kobj_attribute *attr,
 176                                   char *buf)
 177 {
 178         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 179 }
 180 static ssize_t pages_to_scan_store(struct kobject *kobj,
 181                                    struct kobj_attribute *attr,
 182                                    const char *buf, size_t count)
 183 {
 184         int err;
 185         unsigned long pages;
 186
 187         err = kstrtoul(buf, 10, &pages);
 188         if (err || !pages || pages > UINT_MAX)
 189                 return -EINVAL;
 190
 191         khugepaged_pages_to_scan = pages;
 192
 193         return count;
 194 }
 195 static struct kobj_attribute pages_to_scan_attr =
 196         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 197                pages_to_scan_store);
 198
 199 static ssize_t pages_collapsed_show(struct kobject *kobj,
 200                                     struct kobj_attribute *attr,
 201                                     char *buf)
 202 {
 203         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 204 }
 205 static struct kobj_attribute pages_collapsed_attr =
 206         __ATTR_RO(pages_collapsed);
 207
 208 static ssize_t full_scans_show(struct kobject *kobj,
 209                                struct kobj_attribute *attr,
 210                                char *buf)
 211 {
 212         return sprintf(buf, "%u\n", khugepaged_full_scans);
 213 }
 214 static struct kobj_attribute full_scans_attr =
 215         __ATTR_RO(full_scans);
 216
 217 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 218                                       struct kobj_attribute *attr, char *buf)
 219 {
 220         return single_hugepage_flag_show(kobj, attr, buf,
 221                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 222 }
 223 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 224                                        struct kobj_attribute *attr,
 225                                        const char *buf, size_t count)
 226 {
 227         return single_hugepage_flag_store(kobj, attr, buf, count,
 228                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 229 }
 230 static struct kobj_attribute khugepaged_defrag_attr =
 231         __ATTR(defrag, 0644, khugepaged_defrag_show,
 232                khugepaged_defrag_store);
 233
 234 /*
 235  * max_ptes_none controls if khugepaged should collapse hugepages over
 236  * any unmapped ptes in turn potentially increasing the memory
 237  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 238  * reduce the available free memory in the system as it
 239  * runs. Increasing max_ptes_none will instead potentially reduce the
 240  * free memory in the system during the khugepaged scan.
 241  */
 242 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 243                                              struct kobj_attribute *attr,
 244                                              char *buf)
 245 {
 246         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 247 }
 248 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 249                                               struct kobj_attribute *attr,
 250                                               const char *buf, size_t count)
 251 {
 252         int err;
 253         unsigned long max_ptes_none;
 254
 255         err = kstrtoul(buf, 10, &max_ptes_none);
 256         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 257                 return -EINVAL;
 258
 259         khugepaged_max_ptes_none = max_ptes_none;
 260
 261         return count;
 262 }
 263 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 264         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 265                khugepaged_max_ptes_none_store);
 266
 267 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
 268                                              struct kobj_attribute *attr,
 269                                              char *buf)
 270 {
 271         return sprintf(buf, "%u\n", khugepaged_max_ptes_swap);
 272 }
 273
 274 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
 275                                               struct kobj_attribute *attr,
 276                                               const char *buf, size_t count)
 277 {
 278         int err;
 279         unsigned long max_ptes_swap;
 280
 281         err  = kstrtoul(buf, 10, &max_ptes_swap);
 282         if (err || max_ptes_swap > HPAGE_PMD_NR-1)
 283                 return -EINVAL;
 284
 285         khugepaged_max_ptes_swap = max_ptes_swap;
 286
 287         return count;
 288 }
 289
 290 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
 291         __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
 292                khugepaged_max_ptes_swap_store);
 293
 294 static struct attribute *khugepaged_attr[] = {
 295         &khugepaged_defrag_attr.attr,
 296         &khugepaged_max_ptes_none_attr.attr,
 297         &pages_to_scan_attr.attr,
 298         &pages_collapsed_attr.attr,
 299         &full_scans_attr.attr,
 300         &scan_sleep_millisecs_attr.attr,
 301         &alloc_sleep_millisecs_attr.attr,
 302         &khugepaged_max_ptes_swap_attr.attr,
 303         NULL,
 304 };
 305
 306 struct attribute_group khugepaged_attr_group = {
 307         .attrs = khugepaged_attr,
 308         .name = "khugepaged",
 309 };
 310 #endif /* CONFIG_SYSFS */
 311
 312 int hugepage_madvise(struct vm_area_struct *vma,
 313                      unsigned long *vm_flags, int advice)
 314 {
 315         switch (advice) {
 316         case MADV_HUGEPAGE:
 317 #ifdef CONFIG_S390
 318                 /*
 319                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
 320                  * can't handle this properly after s390_enable_sie, so we simply
 321                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
 322                  */
 323                 if (mm_has_pgste(vma->vm_mm))
 324                         return 0;
 325 #endif
 326                 *vm_flags &= ~VM_NOHUGEPAGE;
 327                 *vm_flags |= VM_HUGEPAGE;
 328                 /*
 329                  * If the vma become good for khugepaged to scan,
 330                  * register it here without waiting a page fault that
 331                  * may not happen any time soon.
 332                  */
 333                 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
 334                                 khugepaged_enter_vma_merge(vma, *vm_flags))
 335                         return -ENOMEM;
 336                 break;
 337         case MADV_NOHUGEPAGE:
 338                 *vm_flags &= ~VM_HUGEPAGE;
 339                 *vm_flags |= VM_NOHUGEPAGE;
 340                 /*
 341                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
 342                  * this vma even if we leave the mm registered in khugepaged if
 343                  * it got registered before VM_NOHUGEPAGE was set.
 344                  */
 345                 break;
 346         }
 347
 348         return 0;
 349 }
 350
 351 int __init khugepaged_init(void)
 352 {
 353         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
 354                                           sizeof(struct mm_slot),
 355                                           __alignof__(struct mm_slot), 0, NULL);
 356         if (!mm_slot_cache)
 357                 return -ENOMEM;
 358
 359         khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
 360         khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
 361         khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
 362
 363         return 0;
 364 }
 365
 366 void __init khugepaged_destroy(void)
 367 {
 368         kmem_cache_destroy(mm_slot_cache);
 369 }
 370
 371 static inline struct mm_slot *alloc_mm_slot(void)
 372 {
 373         if (!mm_slot_cache)     /* initialization failed */
 374                 return NULL;
 375         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
 376 }
 377
 378 static inline void free_mm_slot(struct mm_slot *mm_slot)
 379 {
 380         kmem_cache_free(mm_slot_cache, mm_slot);
 381 }
 382
 383 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
 384 {
 385         struct mm_slot *mm_slot;
 386
 387         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
 388                 if (mm == mm_slot->mm)
 389                         return mm_slot;
 390
 391         return NULL;
 392 }
 393
 394 static void insert_to_mm_slots_hash(struct mm_struct *mm,
 395                                     struct mm_slot *mm_slot)
 396 {
 397         mm_slot->mm = mm;
 398         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
 399 }
 400
 401 static inline int khugepaged_test_exit(struct mm_struct *mm)
 402 {
 403         return atomic_read(&mm->mm_users) == 0;
 404 }
 405
 406 static bool hugepage_vma_check(struct vm_area_struct *vma,
 407                                unsigned long vm_flags)
 408 {
 409         if ((!(vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
 410             (vm_flags & VM_NOHUGEPAGE) ||
 411             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
 412                 return false;
 413
 414         if (shmem_file(vma->vm_file) ||
 415             (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) &&
 416              vma->vm_file &&
 417              (vm_flags & VM_DENYWRITE))) {
 418                 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
 419                                 HPAGE_PMD_NR);
 420         }
 421         if (!vma->anon_vma || vma->vm_ops)
 422                 return false;
 423         if (vma_is_temporary_stack(vma))
 424                 return false;
 425         return !(vm_flags & VM_NO_KHUGEPAGED);
 426 }
 427
 428 int __khugepaged_enter(struct mm_struct *mm)
 429 {
 430         struct mm_slot *mm_slot;
 431         int wakeup;
 432
 433         mm_slot = alloc_mm_slot();
 434         if (!mm_slot)
 435                 return -ENOMEM;
 436
 437         /* __khugepaged_exit() must not run from under us */
 438         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
 439         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
 440                 free_mm_slot(mm_slot);
 441                 return 0;
 442         }
 443
 444         spin_lock(&khugepaged_mm_lock);
 445         insert_to_mm_slots_hash(mm, mm_slot);
 446         /*
 447          * Insert just behind the scanning cursor, to let the area settle
 448          * down a little.
 449          */
 450         wakeup = list_empty(&khugepaged_scan.mm_head);
 451         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
 452         spin_unlock(&khugepaged_mm_lock);
 453
 454         mmgrab(mm);
 455         if (wakeup)
 456                 wake_up_interruptible(&khugepaged_wait);
 457
 458         return 0;
 459 }
 460
 461 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
 462                                unsigned long vm_flags)
 463 {
 464         unsigned long hstart, hend;
 465
 466         /*
 467          * khugepaged only supports read-only files for non-shmem files.
 468          * khugepaged does not yet work on special mappings. And
 469          * file-private shmem THP is not supported.
 470          */
 471         if (!hugepage_vma_check(vma, vm_flags))
 472                 return 0;
 473
 474         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 475         hend = vma->vm_end & HPAGE_PMD_MASK;
 476         if (hstart < hend)
 477                 return khugepaged_enter(vma, vm_flags);
 478         return 0;
 479 }
 480
 481 void __khugepaged_exit(struct mm_struct *mm)
 482 {
 483         struct mm_slot *mm_slot;
 484         int free = 0;
 485
 486         spin_lock(&khugepaged_mm_lock);
 487         mm_slot = get_mm_slot(mm);
 488         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
 489                 hash_del(&mm_slot->hash);
 490                 list_del(&mm_slot->mm_node);
 491                 free = 1;
 492         }
 493         spin_unlock(&khugepaged_mm_lock);
 494
 495         if (free) {
 496                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
 497                 free_mm_slot(mm_slot);
 498                 mmdrop(mm);
 499         } else if (mm_slot) {
 500                 /*
 501                  * This is required to serialize against
 502                  * khugepaged_test_exit() (which is guaranteed to run
 503                  * under mmap sem read mode). Stop here (after we
 504                  * return all pagetables will be destroyed) until
 505                  * khugepaged has finished working on the pagetables
 506                  * under the mmap_sem.
 507                  */
 508                 down_write(&mm->mmap_sem);
 509                 up_write(&mm->mmap_sem);
 510         }
 511 }
 512
 513 static void release_pte_page(struct page *page)
 514 {
 515         mod_node_page_state(page_pgdat(page),
 516                         NR_ISOLATED_ANON + page_is_file_lru(page),
 517                         -compound_nr(page));
 518         unlock_page(page);
 519         putback_lru_page(page);
 520 }
 521
 522 static void release_pte_pages(pte_t *pte, pte_t *_pte,
 523                 struct list_head *compound_pagelist)
 524 {
 525         struct page *page, *tmp;
 526
 527         while (--_pte >= pte) {
 528                 pte_t pteval = *_pte;
 529
 530                 page = pte_page(pteval);
 531                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) &&
 532                                 !PageCompound(page))
 533                         release_pte_page(page);
 534         }
 535
 536         list_for_each_entry_safe(page, tmp, compound_pagelist, lru) {
 537                 list_del(&page->lru);
 538                 release_pte_page(page);
 539         }
 540 }
 541
 542 static bool is_refcount_suitable(struct page *page)
 543 {
 544         int expected_refcount;
 545
 546         expected_refcount = total_mapcount(page);
 547         if (PageSwapCache(page))
 548                 expected_refcount += compound_nr(page);
 549
 550         return page_count(page) == expected_refcount;
 551 }
 552
 553 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
 554                                         unsigned long address,
 555                                         pte_t *pte,
 556                                         struct list_head *compound_pagelist)
 557 {
 558         struct page *page = NULL;
 559         pte_t *_pte;
 560         int none_or_zero = 0, result = 0, referenced = 0;
 561         bool writable = false;
 562
 563         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
 564              _pte++, address += PAGE_SIZE) {
 565                 pte_t pteval = *_pte;
 566                 if (pte_none(pteval) || (pte_present(pteval) &&
 567                                 is_zero_pfn(pte_pfn(pteval)))) {
 568                         if (!userfaultfd_armed(vma) &&
 569                             ++none_or_zero <= khugepaged_max_ptes_none) {
 570                                 continue;
 571                         } else {
 572                                 result = SCAN_EXCEED_NONE_PTE;
 573                                 goto out;
 574                         }
 575                 }
 576                 if (!pte_present(pteval)) {
 577                         result = SCAN_PTE_NON_PRESENT;
 578                         goto out;
 579                 }
 580                 page = vm_normal_page(vma, address, pteval);
 581                 if (unlikely(!page)) {
 582                         result = SCAN_PAGE_NULL;
 583                         goto out;
 584                 }
 585
 586                 VM_BUG_ON_PAGE(!PageAnon(page), page);
 587
 588                 if (PageCompound(page)) {
 589                         struct page *p;
 590                         page = compound_head(page);
 591
 592                         /*
 593                          * Check if we have dealt with the compound page
 594                          * already
 595                          */
 596                         list_for_each_entry(p, compound_pagelist, lru) {
 597                                 if (page == p)
 598                                         goto next;
 599                         }
 600                 }
 601
 602                 /*
 603                  * We can do it before isolate_lru_page because the
 604                  * page can't be freed from under us. NOTE: PG_lock
 605                  * is needed to serialize against split_huge_page
 606                  * when invoked from the VM.
 607                  */
 608                 if (!trylock_page(page)) {
 609                         result = SCAN_PAGE_LOCK;
 610                         goto out;
 611                 }
 612
 613                 /*
 614                  * Check if the page has any GUP (or other external) pins.
 615                  *
 616                  * The page table that maps the page has been already unlinked
 617                  * from the page table tree and this process cannot get
 618                  * an additinal pin on the page.
 619                  *
 620                  * New pins can come later if the page is shared across fork,
 621                  * but not from this process. The other process cannot write to
 622                  * the page, only trigger CoW.
 623                  */
 624                 if (!is_refcount_suitable(page)) {
 625                         unlock_page(page);
 626                         result = SCAN_PAGE_COUNT;
 627                         goto out;
 628                 }
 629                 if (!pte_write(pteval) && PageSwapCache(page) &&
 630                                 !reuse_swap_page(page, NULL)) {
 631                         /*
 632                          * Page is in the swap cache and cannot be re-used.
 633                          * It cannot be collapsed into a THP.
 634                          */
 635                         unlock_page(page);
 636                         result = SCAN_SWAP_CACHE_PAGE;
 637                         goto out;
 638                 }
 639
 640                 /*
 641                  * Isolate the page to avoid collapsing an hugepage
 642                  * currently in use by the VM.
 643                  */
 644                 if (isolate_lru_page(page)) {
 645                         unlock_page(page);
 646                         result = SCAN_DEL_PAGE_LRU;
 647                         goto out;
 648                 }
 649                 mod_node_page_state(page_pgdat(page),
 650                                 NR_ISOLATED_ANON + page_is_file_lru(page),
 651                                 compound_nr(page));
 652                 VM_BUG_ON_PAGE(!PageLocked(page), page);
 653                 VM_BUG_ON_PAGE(PageLRU(page), page);
 654
 655                 if (PageCompound(page))
 656                         list_add_tail(&page->lru, compound_pagelist);
 657 next:
 658                 /* There should be enough young pte to collapse the page */
 659                 if (pte_young(pteval) ||
 660                     page_is_young(page) || PageReferenced(page) ||
 661                     mmu_notifier_test_young(vma->vm_mm, address))
 662                         referenced++;
 663
 664                 if (pte_write(pteval))
 665                         writable = true;
 666         }
 667         if (likely(writable)) {
 668                 if (likely(referenced)) {
 669                         result = SCAN_SUCCEED;
 670                         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 671                                                             referenced, writable, result);
 672                         return 1;
 673                 }
 674         } else {
 675                 result = SCAN_PAGE_RO;
 676         }
 677
 678 out:
 679         release_pte_pages(pte, _pte, compound_pagelist);
 680         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 681                                             referenced, writable, result);
 682         return 0;
 683 }
 684
 685 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
 686                                       struct vm_area_struct *vma,
 687                                       unsigned long address,
 688                                       spinlock_t *ptl,
 689                                       struct list_head *compound_pagelist)
 690 {
 691         struct page *src_page, *tmp;
 692         pte_t *_pte;
 693         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 694                                 _pte++, page++, address += PAGE_SIZE) {
 695                 pte_t pteval = *_pte;
 696
 697                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
 698                         clear_user_highpage(page, address);
 699                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
 700                         if (is_zero_pfn(pte_pfn(pteval))) {
 701                                 /*
 702                                  * ptl mostly unnecessary.
 703                                  */
 704                                 spin_lock(ptl);
 705                                 /*
 706                                  * paravirt calls inside pte_clear here are
 707                                  * superfluous.
 708                                  */
 709                                 pte_clear(vma->vm_mm, address, _pte);
 710                                 spin_unlock(ptl);
 711                         }
 712                 } else {
 713                         src_page = pte_page(pteval);
 714                         copy_user_highpage(page, src_page, address, vma);
 715                         if (!PageCompound(src_page))
 716                                 release_pte_page(src_page);
 717                         /*
 718                          * ptl mostly unnecessary, but preempt has to
 719                          * be disabled to update the per-cpu stats
 720                          * inside page_remove_rmap().
 721                          */
 722                         spin_lock(ptl);
 723                         /*
 724                          * paravirt calls inside pte_clear here are
 725                          * superfluous.
 726                          */
 727                         pte_clear(vma->vm_mm, address, _pte);
 728                         page_remove_rmap(src_page, false);
 729                         spin_unlock(ptl);
 730                         free_page_and_swap_cache(src_page);
 731                 }
 732         }
 733
 734         list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
 735                 list_del(&src_page->lru);
 736                 release_pte_page(src_page);
 737         }
 738 }
 739
 740 static void khugepaged_alloc_sleep(void)
 741 {
 742         DEFINE_WAIT(wait);
 743
 744         add_wait_queue(&khugepaged_wait, &wait);
 745         freezable_schedule_timeout_interruptible(
 746                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
 747         remove_wait_queue(&khugepaged_wait, &wait);
 748 }
 749
 750 static int khugepaged_node_load[MAX_NUMNODES];
 751
 752 static bool khugepaged_scan_abort(int nid)
 753 {
 754         int i;
 755
 756         /*
 757          * If node_reclaim_mode is disabled, then no extra effort is made to
 758          * allocate memory locally.
 759          */
 760         if (!node_reclaim_mode)
 761                 return false;
 762
 763         /* If there is a count for this node already, it must be acceptable */
 764         if (khugepaged_node_load[nid])
 765                 return false;
 766
 767         for (i = 0; i < MAX_NUMNODES; i++) {
 768                 if (!khugepaged_node_load[i])
 769                         continue;
 770                 if (node_distance(nid, i) > node_reclaim_distance)
 771                         return true;
 772         }
 773         return false;
 774 }
 775
 776 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
 777 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
 778 {
 779         return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
 780 }
 781
 782 #ifdef CONFIG_NUMA
 783 static int khugepaged_find_target_node(void)
 784 {
 785         static int last_khugepaged_target_node = NUMA_NO_NODE;
 786         int nid, target_node = 0, max_value = 0;
 787
 788         /* find first node with max normal pages hit */
 789         for (nid = 0; nid < MAX_NUMNODES; nid++)
 790                 if (khugepaged_node_load[nid] > max_value) {
 791                         max_value = khugepaged_node_load[nid];
 792                         target_node = nid;
 793                 }
 794
 795         /* do some balance if several nodes have the same hit record */
 796         if (target_node <= last_khugepaged_target_node)
 797                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
 798                                 nid++)
 799                         if (max_value == khugepaged_node_load[nid]) {
 800                                 target_node = nid;
 801                                 break;
 802                         }
 803
 804         last_khugepaged_target_node = target_node;
 805         return target_node;
 806 }
 807
 808 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 809 {
 810         if (IS_ERR(*hpage)) {
 811                 if (!*wait)
 812                         return false;
 813
 814                 *wait = false;
 815                 *hpage = NULL;
 816                 khugepaged_alloc_sleep();
 817         } else if (*hpage) {
 818                 put_page(*hpage);
 819                 *hpage = NULL;
 820         }
 821
 822         return true;
 823 }
 824
 825 static struct page *
 826 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 827 {
 828         VM_BUG_ON_PAGE(*hpage, *hpage);
 829
 830         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
 831         if (unlikely(!*hpage)) {
 832                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 833                 *hpage = ERR_PTR(-ENOMEM);
 834                 return NULL;
 835         }
 836
 837         prep_transhuge_page(*hpage);
 838         count_vm_event(THP_COLLAPSE_ALLOC);
 839         return *hpage;
 840 }
 841 #else
 842 static int khugepaged_find_target_node(void)
 843 {
 844         return 0;
 845 }
 846
 847 static inline struct page *alloc_khugepaged_hugepage(void)
 848 {
 849         struct page *page;
 850
 851         page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
 852                            HPAGE_PMD_ORDER);
 853         if (page)
 854                 prep_transhuge_page(page);
 855         return page;
 856 }
 857
 858 static struct page *khugepaged_alloc_hugepage(bool *wait)
 859 {
 860         struct page *hpage;
 861
 862         do {
 863                 hpage = alloc_khugepaged_hugepage();
 864                 if (!hpage) {
 865                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 866                         if (!*wait)
 867                                 return NULL;
 868
 869                         *wait = false;
 870                         khugepaged_alloc_sleep();
 871                 } else
 872                         count_vm_event(THP_COLLAPSE_ALLOC);
 873         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
 874
 875         return hpage;
 876 }
 877
 878 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 879 {
 880         if (!*hpage)
 881                 *hpage = khugepaged_alloc_hugepage(wait);
 882
 883         if (unlikely(!*hpage))
 884                 return false;
 885
 886         return true;
 887 }
 888
 889 static struct page *
 890 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 891 {
 892         VM_BUG_ON(!*hpage);
 893
 894         return  *hpage;
 895 }
 896 #endif
 897
 898 /*
 899  * If mmap_sem temporarily dropped, revalidate vma
 900  * before taking mmap_sem.
 901  * Return 0 if succeeds, otherwise return none-zero
 902  * value (scan code).
 903  */
 904
 905 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
 906                 struct vm_area_struct **vmap)
 907 {
 908         struct vm_area_struct *vma;
 909         unsigned long hstart, hend;
 910
 911         if (unlikely(khugepaged_test_exit(mm)))
 912                 return SCAN_ANY_PROCESS;
 913
 914         *vmap = vma = find_vma(mm, address);
 915         if (!vma)
 916                 return SCAN_VMA_NULL;
 917
 918         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 919         hend = vma->vm_end & HPAGE_PMD_MASK;
 920         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
 921                 return SCAN_ADDRESS_RANGE;
 922         if (!hugepage_vma_check(vma, vma->vm_flags))
 923                 return SCAN_VMA_CHECK;
 924         return 0;
 925 }
 926
 927 /*
 928  * Bring missing pages in from swap, to complete THP collapse.
 929  * Only done if khugepaged_scan_pmd believes it is worthwhile.
 930  *
 931  * Called and returns without pte mapped or spinlocks held,
 932  * but with mmap_sem held to protect against vma changes.
 933  */
 934
 935 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
 936                                         struct vm_area_struct *vma,
 937                                         unsigned long address, pmd_t *pmd,
 938                                         int referenced)
 939 {
 940         int swapped_in = 0;
 941         vm_fault_t ret = 0;
 942         struct vm_fault vmf = {
 943                 .vma = vma,
 944                 .address = address,
 945                 .flags = FAULT_FLAG_ALLOW_RETRY,
 946                 .pmd = pmd,
 947                 .pgoff = linear_page_index(vma, address),
 948         };
 949
 950         vmf.pte = pte_offset_map(pmd, address);
 951         for (; vmf.address < address + HPAGE_PMD_NR*PAGE_SIZE;
 952                         vmf.pte++, vmf.address += PAGE_SIZE) {
 953                 vmf.orig_pte = *vmf.pte;
 954                 if (!is_swap_pte(vmf.orig_pte))
 955                         continue;
 956                 swapped_in++;
 957                 ret = do_swap_page(&vmf);
 958
 959                 /* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */
 960                 if (ret & VM_FAULT_RETRY) {
 961                         down_read(&mm->mmap_sem);
 962                         if (hugepage_vma_revalidate(mm, address, &vmf.vma)) {
 963                                 /* vma is no longer available, don't continue to swapin */
 964                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 965                                 return false;
 966                         }
 967                         /* check if the pmd is still valid */
 968                         if (mm_find_pmd(mm, address) != pmd) {
 969                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 970                                 return false;
 971                         }
 972                 }
 973                 if (ret & VM_FAULT_ERROR) {
 974                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 975                         return false;
 976                 }
 977                 /* pte is unmapped now, we need to map it */
 978                 vmf.pte = pte_offset_map(pmd, vmf.address);
 979         }
 980         vmf.pte--;
 981         pte_unmap(vmf.pte);
 982
 983         /* Drain LRU add pagevec to remove extra pin on the swapped in pages */
 984         if (swapped_in)
 985                 lru_add_drain();
 986
 987         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
 988         return true;
 989 }
 990
 991 static void collapse_huge_page(struct mm_struct *mm,
 992                                    unsigned long address,
 993                                    struct page **hpage,
 994                                    int node, int referenced, int unmapped)
 995 {
 996         LIST_HEAD(compound_pagelist);
 997         pmd_t *pmd, _pmd;
 998         pte_t *pte;
 999         pgtable_t pgtable;
1000         struct page *new_page;
1001         spinlock_t *pmd_ptl, *pte_ptl;
1002         int isolated = 0, result = 0;
1003         struct mem_cgroup *memcg;
1004         struct vm_area_struct *vma;
1005         struct mmu_notifier_range range;
1006         gfp_t gfp;
1007
1008         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1009
1010         /* Only allocate from the target node */
1011         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1012
1013         /*
1014          * Before allocating the hugepage, release the mmap_sem read lock.
1015          * The allocation can take potentially a long time if it involves
1016          * sync compaction, and we do not need to hold the mmap_sem during
1017          * that. We will recheck the vma after taking it again in write mode.
1018          */
1019         up_read(&mm->mmap_sem);
1020         new_page = khugepaged_alloc_page(hpage, gfp, node);
1021         if (!new_page) {
1022                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1023                 goto out_nolock;
1024         }
1025
1026         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
1027                 result = SCAN_CGROUP_CHARGE_FAIL;
1028                 goto out_nolock;
1029         }
1030
1031         down_read(&mm->mmap_sem);
1032         result = hugepage_vma_revalidate(mm, address, &vma);
1033         if (result) {
1034                 mem_cgroup_cancel_charge(new_page, memcg, true);
1035                 up_read(&mm->mmap_sem);
1036                 goto out_nolock;
1037         }
1038
1039         pmd = mm_find_pmd(mm, address);
1040         if (!pmd) {
1041                 result = SCAN_PMD_NULL;
1042                 mem_cgroup_cancel_charge(new_page, memcg, true);
1043                 up_read(&mm->mmap_sem);
1044                 goto out_nolock;
1045         }
1046
1047         /*
1048          * __collapse_huge_page_swapin always returns with mmap_sem locked.
1049          * If it fails, we release mmap_sem and jump out_nolock.
1050          * Continuing to collapse causes inconsistency.
1051          */
1052         if (unmapped && !__collapse_huge_page_swapin(mm, vma, address,
1053                                                      pmd, referenced)) {
1054                 mem_cgroup_cancel_charge(new_page, memcg, true);
1055                 up_read(&mm->mmap_sem);
1056                 goto out_nolock;
1057         }
1058
1059         up_read(&mm->mmap_sem);
1060         /*
1061          * Prevent all access to pagetables with the exception of
1062          * gup_fast later handled by the ptep_clear_flush and the VM
1063          * handled by the anon_vma lock + PG_lock.
1064          */
1065         down_write(&mm->mmap_sem);
1066         result = SCAN_ANY_PROCESS;
1067         if (!mmget_still_valid(mm))
1068                 goto out;
1069         result = hugepage_vma_revalidate(mm, address, &vma);
1070         if (result)
1071                 goto out;
1072         /* check if the pmd is still valid */
1073         if (mm_find_pmd(mm, address) != pmd)
1074                 goto out;
1075
1076         anon_vma_lock_write(vma->anon_vma);
1077
1078         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1079                                 address, address + HPAGE_PMD_SIZE);
1080         mmu_notifier_invalidate_range_start(&range);
1081
1082         pte = pte_offset_map(pmd, address);
1083         pte_ptl = pte_lockptr(mm, pmd);
1084
1085         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1086         /*
1087          * After this gup_fast can't run anymore. This also removes
1088          * any huge TLB entry from the CPU so we won't allow
1089          * huge and small TLB entries for the same virtual address
1090          * to avoid the risk of CPU bugs in that area.
1091          */
1092         _pmd = pmdp_collapse_flush(vma, address, pmd);
1093         spin_unlock(pmd_ptl);
1094         mmu_notifier_invalidate_range_end(&range);
1095
1096         spin_lock(pte_ptl);
1097         isolated = __collapse_huge_page_isolate(vma, address, pte,
1098                         &compound_pagelist);
1099         spin_unlock(pte_ptl);
1100
1101         if (unlikely(!isolated)) {
1102                 pte_unmap(pte);
1103                 spin_lock(pmd_ptl);
1104                 BUG_ON(!pmd_none(*pmd));
1105                 /*
1106                  * We can only use set_pmd_at when establishing
1107                  * hugepmds and never for establishing regular pmds that
1108                  * points to regular pagetables. Use pmd_populate for that
1109                  */
1110                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1111                 spin_unlock(pmd_ptl);
1112                 anon_vma_unlock_write(vma->anon_vma);
1113                 result = SCAN_FAIL;
1114                 goto out;
1115         }
1116
1117         /*
1118          * All pages are isolated and locked so anon_vma rmap
1119          * can't run anymore.
1120          */
1121         anon_vma_unlock_write(vma->anon_vma);
1122
1123         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl,
1124                         &compound_pagelist);
1125         pte_unmap(pte);
1126         __SetPageUptodate(new_page);
1127         pgtable = pmd_pgtable(_pmd);
1128
1129         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1130         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1131
1132         /*
1133          * spin_lock() below is not the equivalent of smp_wmb(), so
1134          * this is needed to avoid the copy_huge_page writes to become
1135          * visible after the set_pmd_at() write.
1136          */
1137         smp_wmb();
1138
1139         spin_lock(pmd_ptl);
1140         BUG_ON(!pmd_none(*pmd));
1141         page_add_new_anon_rmap(new_page, vma, address, true);
1142         mem_cgroup_commit_charge(new_page, memcg, false, true);
1143         count_memcg_events(memcg, THP_COLLAPSE_ALLOC, 1);
1144         lru_cache_add_active_or_unevictable(new_page, vma);
1145         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1146         set_pmd_at(mm, address, pmd, _pmd);
1147         update_mmu_cache_pmd(vma, address, pmd);
1148         spin_unlock(pmd_ptl);
1149
1150         *hpage = NULL;
1151
1152         khugepaged_pages_collapsed++;
1153         result = SCAN_SUCCEED;
1154 out_up_write:
1155         up_write(&mm->mmap_sem);
1156 out_nolock:
1157         trace_mm_collapse_huge_page(mm, isolated, result);
1158         return;
1159 out:
1160         mem_cgroup_cancel_charge(new_page, memcg, true);
1161         goto out_up_write;
1162 }
1163
1164 static int khugepaged_scan_pmd(struct mm_struct *mm,
1165                                struct vm_area_struct *vma,
1166                                unsigned long address,
1167                                struct page **hpage)
1168 {
1169         pmd_t *pmd;
1170         pte_t *pte, *_pte;
1171         int ret = 0, none_or_zero = 0, result = 0, referenced = 0;
1172         struct page *page = NULL;
1173         unsigned long _address;
1174         spinlock_t *ptl;
1175         int node = NUMA_NO_NODE, unmapped = 0;
1176         bool writable = false;
1177
1178         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1179
1180         pmd = mm_find_pmd(mm, address);
1181         if (!pmd) {
1182                 result = SCAN_PMD_NULL;
1183                 goto out;
1184         }
1185
1186         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1187         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1188         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1189              _pte++, _address += PAGE_SIZE) {
1190                 pte_t pteval = *_pte;
1191                 if (is_swap_pte(pteval)) {
1192                         if (++unmapped <= khugepaged_max_ptes_swap) {
1193                                 /*
1194                                  * Always be strict with uffd-wp
1195                                  * enabled swap entries.  Please see
1196                                  * comment below for pte_uffd_wp().
1197                                  */
1198                                 if (pte_swp_uffd_wp(pteval)) {
1199                                         result = SCAN_PTE_UFFD_WP;
1200                                         goto out_unmap;
1201                                 }
1202                                 continue;
1203                         } else {
1204                                 result = SCAN_EXCEED_SWAP_PTE;
1205                                 goto out_unmap;
1206                         }
1207                 }
1208                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1209                         if (!userfaultfd_armed(vma) &&
1210                             ++none_or_zero <= khugepaged_max_ptes_none) {
1211                                 continue;
1212                         } else {
1213                                 result = SCAN_EXCEED_NONE_PTE;
1214                                 goto out_unmap;
1215                         }
1216                 }
1217                 if (!pte_present(pteval)) {
1218                         result = SCAN_PTE_NON_PRESENT;
1219                         goto out_unmap;
1220                 }
1221                 if (pte_uffd_wp(pteval)) {
1222                         /*
1223                          * Don't collapse the page if any of the small
1224                          * PTEs are armed with uffd write protection.
1225                          * Here we can also mark the new huge pmd as
1226                          * write protected if any of the small ones is
1227                          * marked but that could bring uknown
1228                          * userfault messages that falls outside of
1229                          * the registered range.  So, just be simple.
1230                          */
1231                         result = SCAN_PTE_UFFD_WP;
1232                         goto out_unmap;
1233                 }
1234                 if (pte_write(pteval))
1235                         writable = true;
1236
1237                 page = vm_normal_page(vma, _address, pteval);
1238                 if (unlikely(!page)) {
1239                         result = SCAN_PAGE_NULL;
1240                         goto out_unmap;
1241                 }
1242
1243                 page = compound_head(page);
1244
1245                 /*
1246                  * Record which node the original page is from and save this
1247                  * information to khugepaged_node_load[].
1248                  * Khupaged will allocate hugepage from the node has the max
1249                  * hit record.
1250                  */
1251                 node = page_to_nid(page);
1252                 if (khugepaged_scan_abort(node)) {
1253                         result = SCAN_SCAN_ABORT;
1254                         goto out_unmap;
1255                 }
1256                 khugepaged_node_load[node]++;
1257                 if (!PageLRU(page)) {
1258                         result = SCAN_PAGE_LRU;
1259                         goto out_unmap;
1260                 }
1261                 if (PageLocked(page)) {
1262                         result = SCAN_PAGE_LOCK;
1263                         goto out_unmap;
1264                 }
1265                 if (!PageAnon(page)) {
1266                         result = SCAN_PAGE_ANON;
1267                         goto out_unmap;
1268                 }
1269
1270                 /*
1271                  * Check if the page has any GUP (or other external) pins.
1272                  *
1273                  * Here the check is racy it may see totmal_mapcount > refcount
1274                  * in some cases.
1275                  * For example, one process with one forked child process.
1276                  * The parent has the PMD split due to MADV_DONTNEED, then
1277                  * the child is trying unmap the whole PMD, but khugepaged
1278                  * may be scanning the parent between the child has
1279                  * PageDoubleMap flag cleared and dec the mapcount.  So
1280                  * khugepaged may see total_mapcount > refcount.
1281                  *
1282                  * But such case is ephemeral we could always retry collapse
1283                  * later.  However it may report false positive if the page
1284                  * has excessive GUP pins (i.e. 512).  Anyway the same check
1285                  * will be done again later the risk seems low.
1286                  */
1287                 if (!is_refcount_suitable(page)) {
1288                         result = SCAN_PAGE_COUNT;
1289                         goto out_unmap;
1290                 }
1291                 if (pte_young(pteval) ||
1292                     page_is_young(page) || PageReferenced(page) ||
1293                     mmu_notifier_test_young(vma->vm_mm, address))
1294                         referenced++;
1295         }
1296         if (!writable) {
1297                 result = SCAN_PAGE_RO;
1298         } else if (!referenced || (unmapped && referenced < HPAGE_PMD_NR/2)) {
1299                 result = SCAN_LACK_REFERENCED_PAGE;
1300         } else {
1301                 result = SCAN_SUCCEED;
1302                 ret = 1;
1303         }
1304 out_unmap:
1305         pte_unmap_unlock(pte, ptl);
1306         if (ret) {
1307                 node = khugepaged_find_target_node();
1308                 /* collapse_huge_page will return with the mmap_sem released */
1309                 collapse_huge_page(mm, address, hpage, node,
1310                                 referenced, unmapped);
1311         }
1312 out:
1313         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1314                                      none_or_zero, result, unmapped);
1315         return ret;
1316 }
1317
1318 static void collect_mm_slot(struct mm_slot *mm_slot)
1319 {
1320         struct mm_struct *mm = mm_slot->mm;
1321
1322         lockdep_assert_held(&khugepaged_mm_lock);
1323
1324         if (khugepaged_test_exit(mm)) {
1325                 /* free mm_slot */
1326                 hash_del(&mm_slot->hash);
1327                 list_del(&mm_slot->mm_node);
1328
1329                 /*
1330                  * Not strictly needed because the mm exited already.
1331                  *
1332                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1333                  */
1334
1335                 /* khugepaged_mm_lock actually not necessary for the below */
1336                 free_mm_slot(mm_slot);
1337                 mmdrop(mm);
1338         }
1339 }
1340
1341 #ifdef CONFIG_SHMEM
1342 /*
1343  * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1344  * khugepaged should try to collapse the page table.
1345  */
1346 static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1347                                          unsigned long addr)
1348 {
1349         struct mm_slot *mm_slot;
1350
1351         VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1352
1353         spin_lock(&khugepaged_mm_lock);
1354         mm_slot = get_mm_slot(mm);
1355         if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1356                 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1357         spin_unlock(&khugepaged_mm_lock);
1358         return 0;
1359 }
1360
1361 /**
1362  * Try to collapse a pte-mapped THP for mm at address haddr.
1363  *
1364  * This function checks whether all the PTEs in the PMD are pointing to the
1365  * right THP. If so, retract the page table so the THP can refault in with
1366  * as pmd-mapped.
1367  */
1368 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1369 {
1370         unsigned long haddr = addr & HPAGE_PMD_MASK;
1371         struct vm_area_struct *vma = find_vma(mm, haddr);
1372         struct page *hpage = NULL;
1373         pte_t *start_pte, *pte;
1374         pmd_t *pmd, _pmd;
1375         spinlock_t *ptl;
1376         int count = 0;
1377         int i;
1378
1379         if (!vma || !vma->vm_file ||
1380             vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE)
1381                 return;
1382
1383         /*
1384          * This vm_flags may not have VM_HUGEPAGE if the page was not
1385          * collapsed by this mm. But we can still collapse if the page is
1386          * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1387          * will not fail the vma for missing VM_HUGEPAGE
1388          */
1389         if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE))
1390                 return;
1391
1392         pmd = mm_find_pmd(mm, haddr);
1393         if (!pmd)
1394                 return;
1395
1396         start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1397
1398         /* step 1: check all mapped PTEs are to the right huge page */
1399         for (i = 0, addr = haddr, pte = start_pte;
1400              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1401                 struct page *page;
1402
1403                 /* empty pte, skip */
1404                 if (pte_none(*pte))
1405                         continue;
1406
1407                 /* page swapped out, abort */
1408                 if (!pte_present(*pte))
1409                         goto abort;
1410
1411                 page = vm_normal_page(vma, addr, *pte);
1412
1413                 if (!page || !PageCompound(page))
1414                         goto abort;
1415
1416                 if (!hpage) {
1417                         hpage = compound_head(page);
1418                         /*
1419                          * The mapping of the THP should not change.
1420                          *
1421                          * Note that uprobe, debugger, or MAP_PRIVATE may
1422                          * change the page table, but the new page will
1423                          * not pass PageCompound() check.
1424                          */
1425                         if (WARN_ON(hpage->mapping != vma->vm_file->f_mapping))
1426                                 goto abort;
1427                 }
1428
1429                 /*
1430                  * Confirm the page maps to the correct subpage.
1431                  *
1432                  * Note that uprobe, debugger, or MAP_PRIVATE may change
1433                  * the page table, but the new page will not pass
1434                  * PageCompound() check.
1435                  */
1436                 if (WARN_ON(hpage + i != page))
1437                         goto abort;
1438                 count++;
1439         }
1440
1441         /* step 2: adjust rmap */
1442         for (i = 0, addr = haddr, pte = start_pte;
1443              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1444                 struct page *page;
1445
1446                 if (pte_none(*pte))
1447                         continue;
1448                 page = vm_normal_page(vma, addr, *pte);
1449                 page_remove_rmap(page, false);
1450         }
1451
1452         pte_unmap_unlock(start_pte, ptl);
1453
1454         /* step 3: set proper refcount and mm_counters. */
1455         if (hpage) {
1456                 page_ref_sub(hpage, count);
1457                 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1458         }
1459
1460         /* step 4: collapse pmd */
1461         ptl = pmd_lock(vma->vm_mm, pmd);
1462         _pmd = pmdp_collapse_flush(vma, addr, pmd);
1463         spin_unlock(ptl);
1464         mm_dec_nr_ptes(mm);
1465         pte_free(mm, pmd_pgtable(_pmd));
1466         return;
1467
1468 abort:
1469         pte_unmap_unlock(start_pte, ptl);
1470 }
1471
1472 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1473 {
1474         struct mm_struct *mm = mm_slot->mm;
1475         int i;
1476
1477         if (likely(mm_slot->nr_pte_mapped_thp == 0))
1478                 return 0;
1479
1480         if (!down_write_trylock(&mm->mmap_sem))
1481                 return -EBUSY;
1482
1483         if (unlikely(khugepaged_test_exit(mm)))
1484                 goto out;
1485
1486         for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1487                 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1488
1489 out:
1490         mm_slot->nr_pte_mapped_thp = 0;
1491         up_write(&mm->mmap_sem);
1492         return 0;
1493 }
1494
1495 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1496 {
1497         struct vm_area_struct *vma;
1498         unsigned long addr;
1499         pmd_t *pmd, _pmd;
1500
1501         i_mmap_lock_write(mapping);
1502         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1503                 /*
1504                  * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1505                  * got written to. These VMAs are likely not worth investing
1506                  * down_write(mmap_sem) as PMD-mapping is likely to be split
1507                  * later.
1508                  *
1509                  * Not that vma->anon_vma check is racy: it can be set up after
1510                  * the check but before we took mmap_sem by the fault path.
1511                  * But page lock would prevent establishing any new ptes of the
1512                  * page, so we are safe.
1513                  *
1514                  * An alternative would be drop the check, but check that page
1515                  * table is clear before calling pmdp_collapse_flush() under
1516                  * ptl. It has higher chance to recover THP for the VMA, but
1517                  * has higher cost too.
1518                  */
1519                 if (vma->anon_vma)
1520                         continue;
1521                 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1522                 if (addr & ~HPAGE_PMD_MASK)
1523                         continue;
1524                 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1525                         continue;
1526                 pmd = mm_find_pmd(vma->vm_mm, addr);
1527                 if (!pmd)
1528                         continue;
1529                 /*
1530                  * We need exclusive mmap_sem to retract page table.
1531                  *
1532                  * We use trylock due to lock inversion: we need to acquire
1533                  * mmap_sem while holding page lock. Fault path does it in
1534                  * reverse order. Trylock is a way to avoid deadlock.
1535                  */
1536                 if (down_write_trylock(&vma->vm_mm->mmap_sem)) {
1537                         spinlock_t *ptl = pmd_lock(vma->vm_mm, pmd);
1538                         /* assume page table is clear */
1539                         _pmd = pmdp_collapse_flush(vma, addr, pmd);
1540                         spin_unlock(ptl);
1541                         up_write(&vma->vm_mm->mmap_sem);
1542                         mm_dec_nr_ptes(vma->vm_mm);
1543                         pte_free(vma->vm_mm, pmd_pgtable(_pmd));
1544                 } else {
1545                         /* Try again later */
1546                         khugepaged_add_pte_mapped_thp(vma->vm_mm, addr);
1547                 }
1548         }
1549         i_mmap_unlock_write(mapping);
1550 }
1551
1552 /**
1553  * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1554  *
1555  * Basic scheme is simple, details are more complex:
1556  *  - allocate and lock a new huge page;
1557  *  - scan page cache replacing old pages with the new one
1558  *    + swap/gup in pages if necessary;
1559  *    + fill in gaps;
1560  *    + keep old pages around in case rollback is required;
1561  *  - if replacing succeeds:
1562  *    + copy data over;
1563  *    + free old pages;
1564  *    + unlock huge page;
1565  *  - if replacing failed;
1566  *    + put all pages back and unfreeze them;
1567  *    + restore gaps in the page cache;
1568  *    + unlock and free huge page;
1569  */
1570 static void collapse_file(struct mm_struct *mm,
1571                 struct file *file, pgoff_t start,
1572                 struct page **hpage, int node)
1573 {
1574         struct address_space *mapping = file->f_mapping;
1575         gfp_t gfp;
1576         struct page *new_page;
1577         struct mem_cgroup *memcg;
1578         pgoff_t index, end = start + HPAGE_PMD_NR;
1579         LIST_HEAD(pagelist);
1580         XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1581         int nr_none = 0, result = SCAN_SUCCEED;
1582         bool is_shmem = shmem_file(file);
1583
1584         VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1585         VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1586
1587         /* Only allocate from the target node */
1588         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1589
1590         new_page = khugepaged_alloc_page(hpage, gfp, node);
1591         if (!new_page) {
1592                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1593                 goto out;
1594         }
1595
1596         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
1597                 result = SCAN_CGROUP_CHARGE_FAIL;
1598                 goto out;
1599         }
1600
1601         /* This will be less messy when we use multi-index entries */
1602         do {
1603                 xas_lock_irq(&xas);
1604                 xas_create_range(&xas);
1605                 if (!xas_error(&xas))
1606                         break;
1607                 xas_unlock_irq(&xas);
1608                 if (!xas_nomem(&xas, GFP_KERNEL)) {
1609                         mem_cgroup_cancel_charge(new_page, memcg, true);
1610                         result = SCAN_FAIL;
1611                         goto out;
1612                 }
1613         } while (1);
1614
1615         __SetPageLocked(new_page);
1616         if (is_shmem)
1617                 __SetPageSwapBacked(new_page);
1618         new_page->index = start;
1619         new_page->mapping = mapping;
1620
1621         /*
1622          * At this point the new_page is locked and not up-to-date.
1623          * It's safe to insert it into the page cache, because nobody would
1624          * be able to map it or use it in another way until we unlock it.
1625          */
1626
1627         xas_set(&xas, start);
1628         for (index = start; index < end; index++) {
1629                 struct page *page = xas_next(&xas);
1630
1631                 VM_BUG_ON(index != xas.xa_index);
1632                 if (is_shmem) {
1633                         if (!page) {
1634                                 /*
1635                                  * Stop if extent has been truncated or
1636                                  * hole-punched, and is now completely
1637                                  * empty.
1638                                  */
1639                                 if (index == start) {
1640                                         if (!xas_next_entry(&xas, end - 1)) {
1641                                                 result = SCAN_TRUNCATED;
1642                                                 goto xa_locked;
1643                                         }
1644                                         xas_set(&xas, index);
1645                                 }
1646                                 if (!shmem_charge(mapping->host, 1)) {
1647                                         result = SCAN_FAIL;
1648                                         goto xa_locked;
1649                                 }
1650                                 xas_store(&xas, new_page);
1651                                 nr_none++;
1652                                 continue;
1653                         }
1654
1655                         if (xa_is_value(page) || !PageUptodate(page)) {
1656                                 xas_unlock_irq(&xas);
1657                                 /* swap in or instantiate fallocated page */
1658                                 if (shmem_getpage(mapping->host, index, &page,
1659                                                   SGP_NOHUGE)) {
1660                                         result = SCAN_FAIL;
1661                                         goto xa_unlocked;
1662                                 }
1663                         } else if (trylock_page(page)) {
1664                                 get_page(page);
1665                                 xas_unlock_irq(&xas);
1666                         } else {
1667                                 result = SCAN_PAGE_LOCK;
1668                                 goto xa_locked;
1669                         }
1670                 } else {        /* !is_shmem */
1671                         if (!page || xa_is_value(page)) {
1672                                 xas_unlock_irq(&xas);
1673                                 page_cache_sync_readahead(mapping, &file->f_ra,
1674                                                           file, index,
1675                                                           PAGE_SIZE);
1676                                 /* drain pagevecs to help isolate_lru_page() */
1677                                 lru_add_drain();
1678                                 page = find_lock_page(mapping, index);
1679                                 if (unlikely(page == NULL)) {
1680                                         result = SCAN_FAIL;
1681                                         goto xa_unlocked;
1682                                 }
1683                         } else if (PageDirty(page)) {
1684                                 /*
1685                                  * khugepaged only works on read-only fd,
1686                                  * so this page is dirty because it hasn't
1687                                  * been flushed since first write. There
1688                                  * won't be new dirty pages.
1689                                  *
1690                                  * Trigger async flush here and hope the
1691                                  * writeback is done when khugepaged
1692                                  * revisits this page.
1693                                  *
1694                                  * This is a one-off situation. We are not
1695                                  * forcing writeback in loop.
1696                                  */
1697                                 xas_unlock_irq(&xas);
1698                                 filemap_flush(mapping);
1699                                 result = SCAN_FAIL;
1700                                 goto xa_unlocked;
1701                         } else if (trylock_page(page)) {
1702                                 get_page(page);
1703                                 xas_unlock_irq(&xas);
1704                         } else {
1705                                 result = SCAN_PAGE_LOCK;
1706                                 goto xa_locked;
1707                         }
1708                 }
1709
1710                 /*
1711                  * The page must be locked, so we can drop the i_pages lock
1712                  * without racing with truncate.
1713                  */
1714                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1715
1716                 /* make sure the page is up to date */
1717                 if (unlikely(!PageUptodate(page))) {
1718                         result = SCAN_FAIL;
1719                         goto out_unlock;
1720                 }
1721
1722                 /*
1723                  * If file was truncated then extended, or hole-punched, before
1724                  * we locked the first page, then a THP might be there already.
1725                  */
1726                 if (PageTransCompound(page)) {
1727                         result = SCAN_PAGE_COMPOUND;
1728                         goto out_unlock;
1729                 }
1730
1731                 if (page_mapping(page) != mapping) {
1732                         result = SCAN_TRUNCATED;
1733                         goto out_unlock;
1734                 }
1735
1736                 if (!is_shmem && PageDirty(page)) {
1737                         /*
1738                          * khugepaged only works on read-only fd, so this
1739                          * page is dirty because it hasn't been flushed
1740                          * since first write.
1741                          */
1742                         result = SCAN_FAIL;
1743                         goto out_unlock;
1744                 }
1745
1746                 if (isolate_lru_page(page)) {
1747                         result = SCAN_DEL_PAGE_LRU;
1748                         goto out_unlock;
1749                 }
1750
1751                 if (page_has_private(page) &&
1752                     !try_to_release_page(page, GFP_KERNEL)) {
1753                         result = SCAN_PAGE_HAS_PRIVATE;
1754                         putback_lru_page(page);
1755                         goto out_unlock;
1756                 }
1757
1758                 if (page_mapped(page))
1759                         unmap_mapping_pages(mapping, index, 1, false);
1760
1761                 xas_lock_irq(&xas);
1762                 xas_set(&xas, index);
1763
1764                 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1765                 VM_BUG_ON_PAGE(page_mapped(page), page);
1766
1767                 /*
1768                  * The page is expected to have page_count() == 3:
1769                  *  - we hold a pin on it;
1770                  *  - one reference from page cache;
1771                  *  - one from isolate_lru_page;
1772                  */
1773                 if (!page_ref_freeze(page, 3)) {
1774                         result = SCAN_PAGE_COUNT;
1775                         xas_unlock_irq(&xas);
1776                         putback_lru_page(page);
1777                         goto out_unlock;
1778                 }
1779
1780                 /*
1781                  * Add the page to the list to be able to undo the collapse if
1782                  * something go wrong.
1783                  */
1784                 list_add_tail(&page->lru, &pagelist);
1785
1786                 /* Finally, replace with the new page. */
1787                 xas_store(&xas, new_page);
1788                 continue;
1789 out_unlock:
1790                 unlock_page(page);
1791                 put_page(page);
1792                 goto xa_unlocked;
1793         }
1794
1795         if (is_shmem)
1796                 __inc_node_page_state(new_page, NR_SHMEM_THPS);
1797         else {
1798                 __inc_node_page_state(new_page, NR_FILE_THPS);
1799                 filemap_nr_thps_inc(mapping);
1800         }
1801
1802         if (nr_none) {
1803                 struct zone *zone = page_zone(new_page);
1804
1805                 __mod_node_page_state(zone->zone_pgdat, NR_FILE_PAGES, nr_none);
1806                 if (is_shmem)
1807                         __mod_node_page_state(zone->zone_pgdat,
1808                                               NR_SHMEM, nr_none);
1809         }
1810
1811 xa_locked:
1812         xas_unlock_irq(&xas);
1813 xa_unlocked:
1814
1815         if (result == SCAN_SUCCEED) {
1816                 struct page *page, *tmp;
1817
1818                 /*
1819                  * Replacing old pages with new one has succeeded, now we
1820                  * need to copy the content and free the old pages.
1821                  */
1822                 index = start;
1823                 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1824                         while (index < page->index) {
1825                                 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1826                                 index++;
1827                         }
1828                         copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1829                                         page);
1830                         list_del(&page->lru);
1831                         page->mapping = NULL;
1832                         page_ref_unfreeze(page, 1);
1833                         ClearPageActive(page);
1834                         ClearPageUnevictable(page);
1835                         unlock_page(page);
1836                         put_page(page);
1837                         index++;
1838                 }
1839                 while (index < end) {
1840                         clear_highpage(new_page + (index % HPAGE_PMD_NR));
1841                         index++;
1842                 }
1843
1844                 SetPageUptodate(new_page);
1845                 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1846                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1847
1848                 if (is_shmem) {
1849                         set_page_dirty(new_page);
1850                         lru_cache_add_anon(new_page);
1851                 } else {
1852                         lru_cache_add_file(new_page);
1853                 }
1854                 count_memcg_events(memcg, THP_COLLAPSE_ALLOC, 1);
1855
1856                 /*
1857                  * Remove pte page tables, so we can re-fault the page as huge.
1858                  */
1859                 retract_page_tables(mapping, start);
1860                 *hpage = NULL;
1861
1862                 khugepaged_pages_collapsed++;
1863         } else {
1864                 struct page *page;
1865
1866                 /* Something went wrong: roll back page cache changes */
1867                 xas_lock_irq(&xas);
1868                 mapping->nrpages -= nr_none;
1869
1870                 if (is_shmem)
1871                         shmem_uncharge(mapping->host, nr_none);
1872
1873                 xas_set(&xas, start);
1874                 xas_for_each(&xas, page, end - 1) {
1875                         page = list_first_entry_or_null(&pagelist,
1876                                         struct page, lru);
1877                         if (!page || xas.xa_index < page->index) {
1878                                 if (!nr_none)
1879                                         break;
1880                                 nr_none--;
1881                                 /* Put holes back where they were */
1882                                 xas_store(&xas, NULL);
1883                                 continue;
1884                         }
1885
1886                         VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1887
1888                         /* Unfreeze the page. */
1889                         list_del(&page->lru);
1890                         page_ref_unfreeze(page, 2);
1891                         xas_store(&xas, page);
1892                         xas_pause(&xas);
1893                         xas_unlock_irq(&xas);
1894                         unlock_page(page);
1895                         putback_lru_page(page);
1896                         xas_lock_irq(&xas);
1897                 }
1898                 VM_BUG_ON(nr_none);
1899                 xas_unlock_irq(&xas);
1900
1901                 mem_cgroup_cancel_charge(new_page, memcg, true);
1902                 new_page->mapping = NULL;
1903         }
1904
1905         unlock_page(new_page);
1906 out:
1907         VM_BUG_ON(!list_empty(&pagelist));
1908         /* TODO: tracepoints */
1909 }
1910
1911 static void khugepaged_scan_file(struct mm_struct *mm,
1912                 struct file *file, pgoff_t start, struct page **hpage)
1913 {
1914         struct page *page = NULL;
1915         struct address_space *mapping = file->f_mapping;
1916         XA_STATE(xas, &mapping->i_pages, start);
1917         int present, swap;
1918         int node = NUMA_NO_NODE;
1919         int result = SCAN_SUCCEED;
1920
1921         present = 0;
1922         swap = 0;
1923         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1924         rcu_read_lock();
1925         xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
1926                 if (xas_retry(&xas, page))
1927                         continue;
1928
1929                 if (xa_is_value(page)) {
1930                         if (++swap > khugepaged_max_ptes_swap) {
1931                                 result = SCAN_EXCEED_SWAP_PTE;
1932                                 break;
1933                         }
1934                         continue;
1935                 }
1936
1937                 if (PageTransCompound(page)) {
1938                         result = SCAN_PAGE_COMPOUND;
1939                         break;
1940                 }
1941
1942                 node = page_to_nid(page);
1943                 if (khugepaged_scan_abort(node)) {
1944                         result = SCAN_SCAN_ABORT;
1945                         break;
1946                 }
1947                 khugepaged_node_load[node]++;
1948
1949                 if (!PageLRU(page)) {
1950                         result = SCAN_PAGE_LRU;
1951                         break;
1952                 }
1953
1954                 if (page_count(page) !=
1955                     1 + page_mapcount(page) + page_has_private(page)) {
1956                         result = SCAN_PAGE_COUNT;
1957                         break;
1958                 }
1959
1960                 /*
1961                  * We probably should check if the page is referenced here, but
1962                  * nobody would transfer pte_young() to PageReferenced() for us.
1963                  * And rmap walk here is just too costly...
1964                  */
1965
1966                 present++;
1967
1968                 if (need_resched()) {
1969                         xas_pause(&xas);
1970                         cond_resched_rcu();
1971                 }
1972         }
1973         rcu_read_unlock();
1974
1975         if (result == SCAN_SUCCEED) {
1976                 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
1977                         result = SCAN_EXCEED_NONE_PTE;
1978                 } else {
1979                         node = khugepaged_find_target_node();
1980                         collapse_file(mm, file, start, hpage, node);
1981                 }
1982         }
1983
1984         /* TODO: tracepoints */
1985 }
1986 #else
1987 static void khugepaged_scan_file(struct mm_struct *mm,
1988                 struct file *file, pgoff_t start, struct page **hpage)
1989 {
1990         BUILD_BUG();
1991 }
1992
1993 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1994 {
1995         return 0;
1996 }
1997 #endif
1998
1999 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2000                                             struct page **hpage)
2001         __releases(&khugepaged_mm_lock)
2002         __acquires(&khugepaged_mm_lock)
2003 {
2004         struct mm_slot *mm_slot;
2005         struct mm_struct *mm;
2006         struct vm_area_struct *vma;
2007         int progress = 0;
2008
2009         VM_BUG_ON(!pages);
2010         lockdep_assert_held(&khugepaged_mm_lock);
2011
2012         if (khugepaged_scan.mm_slot)
2013                 mm_slot = khugepaged_scan.mm_slot;
2014         else {
2015                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2016                                      struct mm_slot, mm_node);
2017                 khugepaged_scan.address = 0;
2018                 khugepaged_scan.mm_slot = mm_slot;
2019         }
2020         spin_unlock(&khugepaged_mm_lock);
2021         khugepaged_collapse_pte_mapped_thps(mm_slot);
2022
2023         mm = mm_slot->mm;
2024         /*
2025          * Don't wait for semaphore (to avoid long wait times).  Just move to
2026          * the next mm on the list.
2027          */
2028         vma = NULL;
2029         if (unlikely(!down_read_trylock(&mm->mmap_sem)))
2030                 goto breakouterloop_mmap_sem;
2031         if (likely(!khugepaged_test_exit(mm)))
2032                 vma = find_vma(mm, khugepaged_scan.address);
2033
2034         progress++;
2035         for (; vma; vma = vma->vm_next) {
2036                 unsigned long hstart, hend;
2037
2038                 cond_resched();
2039                 if (unlikely(khugepaged_test_exit(mm))) {
2040                         progress++;
2041                         break;
2042                 }
2043                 if (!hugepage_vma_check(vma, vma->vm_flags)) {
2044 skip:
2045                         progress++;
2046                         continue;
2047                 }
2048                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2049                 hend = vma->vm_end & HPAGE_PMD_MASK;
2050                 if (hstart >= hend)
2051                         goto skip;
2052                 if (khugepaged_scan.address > hend)
2053                         goto skip;
2054                 if (khugepaged_scan.address < hstart)
2055                         khugepaged_scan.address = hstart;
2056                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2057                 if (shmem_file(vma->vm_file) && !shmem_huge_enabled(vma))
2058                         goto skip;
2059
2060                 while (khugepaged_scan.address < hend) {
2061                         int ret;
2062                         cond_resched();
2063                         if (unlikely(khugepaged_test_exit(mm)))
2064                                 goto breakouterloop;
2065
2066                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2067                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2068                                   hend);
2069                         if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2070                                 struct file *file = get_file(vma->vm_file);
2071                                 pgoff_t pgoff = linear_page_index(vma,
2072                                                 khugepaged_scan.address);
2073
2074                                 up_read(&mm->mmap_sem);
2075                                 ret = 1;
2076                                 khugepaged_scan_file(mm, file, pgoff, hpage);
2077                                 fput(file);
2078                         } else {
2079                                 ret = khugepaged_scan_pmd(mm, vma,
2080                                                 khugepaged_scan.address,
2081                                                 hpage);
2082                         }
2083                         /* move to next address */
2084                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2085                         progress += HPAGE_PMD_NR;
2086                         if (ret)
2087                                 /* we released mmap_sem so break loop */
2088                                 goto breakouterloop_mmap_sem;
2089                         if (progress >= pages)
2090                                 goto breakouterloop;
2091                 }
2092         }
2093 breakouterloop:
2094         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2095 breakouterloop_mmap_sem:
2096
2097         spin_lock(&khugepaged_mm_lock);
2098         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2099         /*
2100          * Release the current mm_slot if this mm is about to die, or
2101          * if we scanned all vmas of this mm.
2102          */
2103         if (khugepaged_test_exit(mm) || !vma) {
2104                 /*
2105                  * Make sure that if mm_users is reaching zero while
2106                  * khugepaged runs here, khugepaged_exit will find
2107                  * mm_slot not pointing to the exiting mm.
2108                  */
2109                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2110                         khugepaged_scan.mm_slot = list_entry(
2111                                 mm_slot->mm_node.next,
2112                                 struct mm_slot, mm_node);
2113                         khugepaged_scan.address = 0;
2114                 } else {
2115                         khugepaged_scan.mm_slot = NULL;
2116                         khugepaged_full_scans++;
2117                 }
2118
2119                 collect_mm_slot(mm_slot);
2120         }
2121
2122         return progress;
2123 }
2124
2125 static int khugepaged_has_work(void)
2126 {
2127         return !list_empty(&khugepaged_scan.mm_head) &&
2128                 khugepaged_enabled();
2129 }
2130
2131 static int khugepaged_wait_event(void)
2132 {
2133         return !list_empty(&khugepaged_scan.mm_head) ||
2134                 kthread_should_stop();
2135 }
2136
2137 static void khugepaged_do_scan(void)
2138 {
2139         struct page *hpage = NULL;
2140         unsigned int progress = 0, pass_through_head = 0;
2141         unsigned int pages = khugepaged_pages_to_scan;
2142         bool wait = true;
2143
2144         barrier(); /* write khugepaged_pages_to_scan to local stack */
2145
2146         lru_add_drain_all();
2147
2148         while (progress < pages) {
2149                 if (!khugepaged_prealloc_page(&hpage, &wait))
2150                         break;
2151
2152                 cond_resched();
2153
2154                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2155                         break;
2156
2157                 spin_lock(&khugepaged_mm_lock);
2158                 if (!khugepaged_scan.mm_slot)
2159                         pass_through_head++;
2160                 if (khugepaged_has_work() &&
2161                     pass_through_head < 2)
2162                         progress += khugepaged_scan_mm_slot(pages - progress,
2163                                                             &hpage);
2164                 else
2165                         progress = pages;
2166                 spin_unlock(&khugepaged_mm_lock);
2167         }
2168
2169         if (!IS_ERR_OR_NULL(hpage))
2170                 put_page(hpage);
2171 }
2172
2173 static bool khugepaged_should_wakeup(void)
2174 {
2175         return kthread_should_stop() ||
2176                time_after_eq(jiffies, khugepaged_sleep_expire);
2177 }
2178
2179 static void khugepaged_wait_work(void)
2180 {
2181         if (khugepaged_has_work()) {
2182                 const unsigned long scan_sleep_jiffies =
2183                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2184
2185                 if (!scan_sleep_jiffies)
2186                         return;
2187
2188                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2189                 wait_event_freezable_timeout(khugepaged_wait,
2190                                              khugepaged_should_wakeup(),
2191                                              scan_sleep_jiffies);
2192                 return;
2193         }
2194
2195         if (khugepaged_enabled())
2196                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2197 }
2198
2199 static int khugepaged(void *none)
2200 {
2201         struct mm_slot *mm_slot;
2202
2203         set_freezable();
2204         set_user_nice(current, MAX_NICE);
2205
2206         while (!kthread_should_stop()) {
2207                 khugepaged_do_scan();
2208                 khugepaged_wait_work();
2209         }
2210
2211         spin_lock(&khugepaged_mm_lock);
2212         mm_slot = khugepaged_scan.mm_slot;
2213         khugepaged_scan.mm_slot = NULL;
2214         if (mm_slot)
2215                 collect_mm_slot(mm_slot);
2216         spin_unlock(&khugepaged_mm_lock);
2217         return 0;
2218 }
2219
2220 static void set_recommended_min_free_kbytes(void)
2221 {
2222         struct zone *zone;
2223         int nr_zones = 0;
2224         unsigned long recommended_min;
2225
2226         for_each_populated_zone(zone) {
2227                 /*
2228                  * We don't need to worry about fragmentation of
2229                  * ZONE_MOVABLE since it only has movable pages.
2230                  */
2231                 if (zone_idx(zone) > gfp_zone(GFP_USER))
2232                         continue;
2233
2234                 nr_zones++;
2235         }
2236
2237         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2238         recommended_min = pageblock_nr_pages * nr_zones * 2;
2239
2240         /*
2241          * Make sure that on average at least two pageblocks are almost free
2242          * of another type, one for a migratetype to fall back to and a
2243          * second to avoid subsequent fallbacks of other types There are 3
2244          * MIGRATE_TYPES we care about.
2245          */
2246         recommended_min += pageblock_nr_pages * nr_zones *
2247                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2248
2249         /* don't ever allow to reserve more than 5% of the lowmem */
2250         recommended_min = min(recommended_min,
2251                               (unsigned long) nr_free_buffer_pages() / 20);
2252         recommended_min <<= (PAGE_SHIFT-10);
2253
2254         if (recommended_min > min_free_kbytes) {
2255                 if (user_min_free_kbytes >= 0)
2256                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2257                                 min_free_kbytes, recommended_min);
2258
2259                 min_free_kbytes = recommended_min;
2260         }
2261         setup_per_zone_wmarks();
2262 }
2263
2264 int start_stop_khugepaged(void)
2265 {
2266         static struct task_struct *khugepaged_thread __read_mostly;
2267         static DEFINE_MUTEX(khugepaged_mutex);
2268         int err = 0;
2269
2270         mutex_lock(&khugepaged_mutex);
2271         if (khugepaged_enabled()) {
2272                 if (!khugepaged_thread)
2273                         khugepaged_thread = kthread_run(khugepaged, NULL,
2274                                                         "khugepaged");
2275                 if (IS_ERR(khugepaged_thread)) {
2276                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2277                         err = PTR_ERR(khugepaged_thread);
2278                         khugepaged_thread = NULL;
2279                         goto fail;
2280                 }
2281
2282                 if (!list_empty(&khugepaged_scan.mm_head))
2283                         wake_up_interruptible(&khugepaged_wait);
2284
2285                 set_recommended_min_free_kbytes();
2286         } else if (khugepaged_thread) {
2287                 kthread_stop(khugepaged_thread);
2288                 khugepaged_thread = NULL;
2289         }
2290 fail:
2291         mutex_unlock(&khugepaged_mutex);
2292         return err;
2293 }