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/page_table_check.h>
  20 #include <linux/rcupdate_wait.h>
  21 #include <linux/swapops.h>
  22 #include <linux/shmem_fs.h>
  23 #include <linux/ksm.h>
  24
  25 #include <asm/tlb.h>
  26 #include <asm/pgalloc.h>
  27 #include "internal.h"
  28 #include "mm_slot.h"
  29
  30 enum scan_result {
  31         SCAN_FAIL,
  32         SCAN_SUCCEED,
  33         SCAN_PMD_NULL,
  34         SCAN_PMD_NONE,
  35         SCAN_PMD_MAPPED,
  36         SCAN_EXCEED_NONE_PTE,
  37         SCAN_EXCEED_SWAP_PTE,
  38         SCAN_EXCEED_SHARED_PTE,
  39         SCAN_PTE_NON_PRESENT,
  40         SCAN_PTE_UFFD_WP,
  41         SCAN_PTE_MAPPED_HUGEPAGE,
  42         SCAN_PAGE_RO,
  43         SCAN_LACK_REFERENCED_PAGE,
  44         SCAN_PAGE_NULL,
  45         SCAN_SCAN_ABORT,
  46         SCAN_PAGE_COUNT,
  47         SCAN_PAGE_LRU,
  48         SCAN_PAGE_LOCK,
  49         SCAN_PAGE_ANON,
  50         SCAN_PAGE_COMPOUND,
  51         SCAN_ANY_PROCESS,
  52         SCAN_VMA_NULL,
  53         SCAN_VMA_CHECK,
  54         SCAN_ADDRESS_RANGE,
  55         SCAN_DEL_PAGE_LRU,
  56         SCAN_ALLOC_HUGE_PAGE_FAIL,
  57         SCAN_CGROUP_CHARGE_FAIL,
  58         SCAN_TRUNCATED,
  59         SCAN_PAGE_HAS_PRIVATE,
  60         SCAN_STORE_FAILED,
  61         SCAN_COPY_MC,
  62         SCAN_PAGE_FILLED,
  63 };
  64
  65 #define CREATE_TRACE_POINTS
  66 #include <trace/events/huge_memory.h>
  67
  68 static struct task_struct *khugepaged_thread __read_mostly;
  69 static DEFINE_MUTEX(khugepaged_mutex);
  70
  71 /* default scan 8*512 pte (or vmas) every 30 second */
  72 static unsigned int khugepaged_pages_to_scan __read_mostly;
  73 static unsigned int khugepaged_pages_collapsed;
  74 static unsigned int khugepaged_full_scans;
  75 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  76 /* during fragmentation poll the hugepage allocator once every minute */
  77 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  78 static unsigned long khugepaged_sleep_expire;
  79 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  80 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  81 /*
  82  * default collapse hugepages if there is at least one pte mapped like
  83  * it would have happened if the vma was large enough during page
  84  * fault.
  85  *
  86  * Note that these are only respected if collapse was initiated by khugepaged.
  87  */
  88 static unsigned int khugepaged_max_ptes_none __read_mostly;
  89 static unsigned int khugepaged_max_ptes_swap __read_mostly;
  90 static unsigned int khugepaged_max_ptes_shared __read_mostly;
  91
  92 #define MM_SLOTS_HASH_BITS 10
  93 static DEFINE_READ_MOSTLY_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  94
  95 static struct kmem_cache *mm_slot_cache __ro_after_init;
  96
  97 struct collapse_control {
  98         bool is_khugepaged;
  99
 100         /* Num pages scanned per node */
 101         u32 node_load[MAX_NUMNODES];
 102
 103         /* nodemask for allocation fallback */
 104         nodemask_t alloc_nmask;
 105 };
 106
 107 /**
 108  * struct khugepaged_mm_slot - khugepaged information per mm that is being scanned
 109  * @slot: hash lookup from mm to mm_slot
 110  */
 111 struct khugepaged_mm_slot {
 112         struct mm_slot slot;
 113 };
 114
 115 /**
 116  * struct khugepaged_scan - cursor for scanning
 117  * @mm_head: the head of the mm list to scan
 118  * @mm_slot: the current mm_slot we are scanning
 119  * @address: the next address inside that to be scanned
 120  *
 121  * There is only the one khugepaged_scan instance of this cursor structure.
 122  */
 123 struct khugepaged_scan {
 124         struct list_head mm_head;
 125         struct khugepaged_mm_slot *mm_slot;
 126         unsigned long address;
 127 };
 128
 129 static struct khugepaged_scan khugepaged_scan = {
 130         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 131 };
 132
 133 #ifdef CONFIG_SYSFS
 134 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 135                                          struct kobj_attribute *attr,
 136                                          char *buf)
 137 {
 138         return sysfs_emit(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 139 }
 140
 141 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 142                                           struct kobj_attribute *attr,
 143                                           const char *buf, size_t count)
 144 {
 145         unsigned int msecs;
 146         int err;
 147
 148         err = kstrtouint(buf, 10, &msecs);
 149         if (err)
 150                 return -EINVAL;
 151
 152         khugepaged_scan_sleep_millisecs = msecs;
 153         khugepaged_sleep_expire = 0;
 154         wake_up_interruptible(&khugepaged_wait);
 155
 156         return count;
 157 }
 158 static struct kobj_attribute scan_sleep_millisecs_attr =
 159         __ATTR_RW(scan_sleep_millisecs);
 160
 161 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 162                                           struct kobj_attribute *attr,
 163                                           char *buf)
 164 {
 165         return sysfs_emit(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 166 }
 167
 168 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 169                                            struct kobj_attribute *attr,
 170                                            const char *buf, size_t count)
 171 {
 172         unsigned int msecs;
 173         int err;
 174
 175         err = kstrtouint(buf, 10, &msecs);
 176         if (err)
 177                 return -EINVAL;
 178
 179         khugepaged_alloc_sleep_millisecs = msecs;
 180         khugepaged_sleep_expire = 0;
 181         wake_up_interruptible(&khugepaged_wait);
 182
 183         return count;
 184 }
 185 static struct kobj_attribute alloc_sleep_millisecs_attr =
 186         __ATTR_RW(alloc_sleep_millisecs);
 187
 188 static ssize_t pages_to_scan_show(struct kobject *kobj,
 189                                   struct kobj_attribute *attr,
 190                                   char *buf)
 191 {
 192         return sysfs_emit(buf, "%u\n", khugepaged_pages_to_scan);
 193 }
 194 static ssize_t pages_to_scan_store(struct kobject *kobj,
 195                                    struct kobj_attribute *attr,
 196                                    const char *buf, size_t count)
 197 {
 198         unsigned int pages;
 199         int err;
 200
 201         err = kstrtouint(buf, 10, &pages);
 202         if (err || !pages)
 203                 return -EINVAL;
 204
 205         khugepaged_pages_to_scan = pages;
 206
 207         return count;
 208 }
 209 static struct kobj_attribute pages_to_scan_attr =
 210         __ATTR_RW(pages_to_scan);
 211
 212 static ssize_t pages_collapsed_show(struct kobject *kobj,
 213                                     struct kobj_attribute *attr,
 214                                     char *buf)
 215 {
 216         return sysfs_emit(buf, "%u\n", khugepaged_pages_collapsed);
 217 }
 218 static struct kobj_attribute pages_collapsed_attr =
 219         __ATTR_RO(pages_collapsed);
 220
 221 static ssize_t full_scans_show(struct kobject *kobj,
 222                                struct kobj_attribute *attr,
 223                                char *buf)
 224 {
 225         return sysfs_emit(buf, "%u\n", khugepaged_full_scans);
 226 }
 227 static struct kobj_attribute full_scans_attr =
 228         __ATTR_RO(full_scans);
 229
 230 static ssize_t defrag_show(struct kobject *kobj,
 231                            struct kobj_attribute *attr, char *buf)
 232 {
 233         return single_hugepage_flag_show(kobj, attr, buf,
 234                                          TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 235 }
 236 static ssize_t defrag_store(struct kobject *kobj,
 237                             struct kobj_attribute *attr,
 238                             const char *buf, size_t count)
 239 {
 240         return single_hugepage_flag_store(kobj, attr, buf, count,
 241                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 242 }
 243 static struct kobj_attribute khugepaged_defrag_attr =
 244         __ATTR_RW(defrag);
 245
 246 /*
 247  * max_ptes_none controls if khugepaged should collapse hugepages over
 248  * any unmapped ptes in turn potentially increasing the memory
 249  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 250  * reduce the available free memory in the system as it
 251  * runs. Increasing max_ptes_none will instead potentially reduce the
 252  * free memory in the system during the khugepaged scan.
 253  */
 254 static ssize_t max_ptes_none_show(struct kobject *kobj,
 255                                   struct kobj_attribute *attr,
 256                                   char *buf)
 257 {
 258         return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_none);
 259 }
 260 static ssize_t max_ptes_none_store(struct kobject *kobj,
 261                                    struct kobj_attribute *attr,
 262                                    const char *buf, size_t count)
 263 {
 264         int err;
 265         unsigned long max_ptes_none;
 266
 267         err = kstrtoul(buf, 10, &max_ptes_none);
 268         if (err || max_ptes_none > HPAGE_PMD_NR - 1)
 269                 return -EINVAL;
 270
 271         khugepaged_max_ptes_none = max_ptes_none;
 272
 273         return count;
 274 }
 275 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 276         __ATTR_RW(max_ptes_none);
 277
 278 static ssize_t max_ptes_swap_show(struct kobject *kobj,
 279                                   struct kobj_attribute *attr,
 280                                   char *buf)
 281 {
 282         return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_swap);
 283 }
 284
 285 static ssize_t max_ptes_swap_store(struct kobject *kobj,
 286                                    struct kobj_attribute *attr,
 287                                    const char *buf, size_t count)
 288 {
 289         int err;
 290         unsigned long max_ptes_swap;
 291
 292         err  = kstrtoul(buf, 10, &max_ptes_swap);
 293         if (err || max_ptes_swap > HPAGE_PMD_NR - 1)
 294                 return -EINVAL;
 295
 296         khugepaged_max_ptes_swap = max_ptes_swap;
 297
 298         return count;
 299 }
 300
 301 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
 302         __ATTR_RW(max_ptes_swap);
 303
 304 static ssize_t max_ptes_shared_show(struct kobject *kobj,
 305                                     struct kobj_attribute *attr,
 306                                     char *buf)
 307 {
 308         return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_shared);
 309 }
 310
 311 static ssize_t max_ptes_shared_store(struct kobject *kobj,
 312                                      struct kobj_attribute *attr,
 313                                      const char *buf, size_t count)
 314 {
 315         int err;
 316         unsigned long max_ptes_shared;
 317
 318         err  = kstrtoul(buf, 10, &max_ptes_shared);
 319         if (err || max_ptes_shared > HPAGE_PMD_NR - 1)
 320                 return -EINVAL;
 321
 322         khugepaged_max_ptes_shared = max_ptes_shared;
 323
 324         return count;
 325 }
 326
 327 static struct kobj_attribute khugepaged_max_ptes_shared_attr =
 328         __ATTR_RW(max_ptes_shared);
 329
 330 static struct attribute *khugepaged_attr[] = {
 331         &khugepaged_defrag_attr.attr,
 332         &khugepaged_max_ptes_none_attr.attr,
 333         &khugepaged_max_ptes_swap_attr.attr,
 334         &khugepaged_max_ptes_shared_attr.attr,
 335         &pages_to_scan_attr.attr,
 336         &pages_collapsed_attr.attr,
 337         &full_scans_attr.attr,
 338         &scan_sleep_millisecs_attr.attr,
 339         &alloc_sleep_millisecs_attr.attr,
 340         NULL,
 341 };
 342
 343 struct attribute_group khugepaged_attr_group = {
 344         .attrs = khugepaged_attr,
 345         .name = "khugepaged",
 346 };
 347 #endif /* CONFIG_SYSFS */
 348
 349 int hugepage_madvise(struct vm_area_struct *vma,
 350                      unsigned long *vm_flags, int advice)
 351 {
 352         switch (advice) {
 353         case MADV_HUGEPAGE:
 354 #ifdef CONFIG_S390
 355                 /*
 356                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
 357                  * can't handle this properly after s390_enable_sie, so we simply
 358                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
 359                  */
 360                 if (mm_has_pgste(vma->vm_mm))
 361                         return 0;
 362 #endif
 363                 *vm_flags &= ~VM_NOHUGEPAGE;
 364                 *vm_flags |= VM_HUGEPAGE;
 365                 /*
 366                  * If the vma become good for khugepaged to scan,
 367                  * register it here without waiting a page fault that
 368                  * may not happen any time soon.
 369                  */
 370                 khugepaged_enter_vma(vma, *vm_flags);
 371                 break;
 372         case MADV_NOHUGEPAGE:
 373                 *vm_flags &= ~VM_HUGEPAGE;
 374                 *vm_flags |= VM_NOHUGEPAGE;
 375                 /*
 376                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
 377                  * this vma even if we leave the mm registered in khugepaged if
 378                  * it got registered before VM_NOHUGEPAGE was set.
 379                  */
 380                 break;
 381         }
 382
 383         return 0;
 384 }
 385
 386 int __init khugepaged_init(void)
 387 {
 388         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
 389                                           sizeof(struct khugepaged_mm_slot),
 390                                           __alignof__(struct khugepaged_mm_slot),
 391                                           0, NULL);
 392         if (!mm_slot_cache)
 393                 return -ENOMEM;
 394
 395         khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
 396         khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
 397         khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
 398         khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2;
 399
 400         return 0;
 401 }
 402
 403 void __init khugepaged_destroy(void)
 404 {
 405         kmem_cache_destroy(mm_slot_cache);
 406 }
 407
 408 static inline int hpage_collapse_test_exit(struct mm_struct *mm)
 409 {
 410         return atomic_read(&mm->mm_users) == 0;
 411 }
 412
 413 void __khugepaged_enter(struct mm_struct *mm)
 414 {
 415         struct khugepaged_mm_slot *mm_slot;
 416         struct mm_slot *slot;
 417         int wakeup;
 418
 419         /* __khugepaged_exit() must not run from under us */
 420         VM_BUG_ON_MM(hpage_collapse_test_exit(mm), mm);
 421         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags)))
 422                 return;
 423
 424         mm_slot = mm_slot_alloc(mm_slot_cache);
 425         if (!mm_slot)
 426                 return;
 427
 428         slot = &mm_slot->slot;
 429
 430         spin_lock(&khugepaged_mm_lock);
 431         mm_slot_insert(mm_slots_hash, mm, slot);
 432         /*
 433          * Insert just behind the scanning cursor, to let the area settle
 434          * down a little.
 435          */
 436         wakeup = list_empty(&khugepaged_scan.mm_head);
 437         list_add_tail(&slot->mm_node, &khugepaged_scan.mm_head);
 438         spin_unlock(&khugepaged_mm_lock);
 439
 440         mmgrab(mm);
 441         if (wakeup)
 442                 wake_up_interruptible(&khugepaged_wait);
 443 }
 444
 445 void khugepaged_enter_vma(struct vm_area_struct *vma,
 446                           unsigned long vm_flags)
 447 {
 448         if (!test_bit(MMF_VM_HUGEPAGE, &vma->vm_mm->flags) &&
 449             hugepage_flags_enabled()) {
 450                 if (thp_vma_allowable_order(vma, vm_flags, false, false, true,
 451                                             PMD_ORDER))
 452                         __khugepaged_enter(vma->vm_mm);
 453         }
 454 }
 455
 456 void __khugepaged_exit(struct mm_struct *mm)
 457 {
 458         struct khugepaged_mm_slot *mm_slot;
 459         struct mm_slot *slot;
 460         int free = 0;
 461
 462         spin_lock(&khugepaged_mm_lock);
 463         slot = mm_slot_lookup(mm_slots_hash, mm);
 464         mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot);
 465         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
 466                 hash_del(&slot->hash);
 467                 list_del(&slot->mm_node);
 468                 free = 1;
 469         }
 470         spin_unlock(&khugepaged_mm_lock);
 471
 472         if (free) {
 473                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
 474                 mm_slot_free(mm_slot_cache, mm_slot);
 475                 mmdrop(mm);
 476         } else if (mm_slot) {
 477                 /*
 478                  * This is required to serialize against
 479                  * hpage_collapse_test_exit() (which is guaranteed to run
 480                  * under mmap sem read mode). Stop here (after we return all
 481                  * pagetables will be destroyed) until khugepaged has finished
 482                  * working on the pagetables under the mmap_lock.
 483                  */
 484                 mmap_write_lock(mm);
 485                 mmap_write_unlock(mm);
 486         }
 487 }
 488
 489 static void release_pte_folio(struct folio *folio)
 490 {
 491         node_stat_mod_folio(folio,
 492                         NR_ISOLATED_ANON + folio_is_file_lru(folio),
 493                         -folio_nr_pages(folio));
 494         folio_unlock(folio);
 495         folio_putback_lru(folio);
 496 }
 497
 498 static void release_pte_pages(pte_t *pte, pte_t *_pte,
 499                 struct list_head *compound_pagelist)
 500 {
 501         struct folio *folio, *tmp;
 502
 503         while (--_pte >= pte) {
 504                 pte_t pteval = ptep_get(_pte);
 505                 unsigned long pfn;
 506
 507                 if (pte_none(pteval))
 508                         continue;
 509                 pfn = pte_pfn(pteval);
 510                 if (is_zero_pfn(pfn))
 511                         continue;
 512                 folio = pfn_folio(pfn);
 513                 if (folio_test_large(folio))
 514                         continue;
 515                 release_pte_folio(folio);
 516         }
 517
 518         list_for_each_entry_safe(folio, tmp, compound_pagelist, lru) {
 519                 list_del(&folio->lru);
 520                 release_pte_folio(folio);
 521         }
 522 }
 523
 524 static bool is_refcount_suitable(struct folio *folio)
 525 {
 526         int expected_refcount;
 527
 528         expected_refcount = folio_mapcount(folio);
 529         if (folio_test_swapcache(folio))
 530                 expected_refcount += folio_nr_pages(folio);
 531
 532         return folio_ref_count(folio) == expected_refcount;
 533 }
 534
 535 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
 536                                         unsigned long address,
 537                                         pte_t *pte,
 538                                         struct collapse_control *cc,
 539                                         struct list_head *compound_pagelist)
 540 {
 541         struct page *page = NULL;
 542         struct folio *folio = NULL;
 543         pte_t *_pte;
 544         int none_or_zero = 0, shared = 0, result = SCAN_FAIL, referenced = 0;
 545         bool writable = false;
 546
 547         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 548              _pte++, address += PAGE_SIZE) {
 549                 pte_t pteval = ptep_get(_pte);
 550                 if (pte_none(pteval) || (pte_present(pteval) &&
 551                                 is_zero_pfn(pte_pfn(pteval)))) {
 552                         ++none_or_zero;
 553                         if (!userfaultfd_armed(vma) &&
 554                             (!cc->is_khugepaged ||
 555                              none_or_zero <= khugepaged_max_ptes_none)) {
 556                                 continue;
 557                         } else {
 558                                 result = SCAN_EXCEED_NONE_PTE;
 559                                 count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
 560                                 goto out;
 561                         }
 562                 }
 563                 if (!pte_present(pteval)) {
 564                         result = SCAN_PTE_NON_PRESENT;
 565                         goto out;
 566                 }
 567                 if (pte_uffd_wp(pteval)) {
 568                         result = SCAN_PTE_UFFD_WP;
 569                         goto out;
 570                 }
 571                 page = vm_normal_page(vma, address, pteval);
 572                 if (unlikely(!page) || unlikely(is_zone_device_page(page))) {
 573                         result = SCAN_PAGE_NULL;
 574                         goto out;
 575                 }
 576
 577                 folio = page_folio(page);
 578                 VM_BUG_ON_FOLIO(!folio_test_anon(folio), folio);
 579
 580                 if (page_mapcount(page) > 1) {
 581                         ++shared;
 582                         if (cc->is_khugepaged &&
 583                             shared > khugepaged_max_ptes_shared) {
 584                                 result = SCAN_EXCEED_SHARED_PTE;
 585                                 count_vm_event(THP_SCAN_EXCEED_SHARED_PTE);
 586                                 goto out;
 587                         }
 588                 }
 589
 590                 if (folio_test_large(folio)) {
 591                         struct folio *f;
 592
 593                         /*
 594                          * Check if we have dealt with the compound page
 595                          * already
 596                          */
 597                         list_for_each_entry(f, compound_pagelist, lru) {
 598                                 if (folio == f)
 599                                         goto next;
 600                         }
 601                 }
 602
 603                 /*
 604                  * We can do it before isolate_lru_page because the
 605                  * page can't be freed from under us. NOTE: PG_lock
 606                  * is needed to serialize against split_huge_page
 607                  * when invoked from the VM.
 608                  */
 609                 if (!folio_trylock(folio)) {
 610                         result = SCAN_PAGE_LOCK;
 611                         goto out;
 612                 }
 613
 614                 /*
 615                  * Check if the page has any GUP (or other external) pins.
 616                  *
 617                  * The page table that maps the page has been already unlinked
 618                  * from the page table tree and this process cannot get
 619                  * an additional pin on the page.
 620                  *
 621                  * New pins can come later if the page is shared across fork,
 622                  * but not from this process. The other process cannot write to
 623                  * the page, only trigger CoW.
 624                  */
 625                 if (!is_refcount_suitable(folio)) {
 626                         folio_unlock(folio);
 627                         result = SCAN_PAGE_COUNT;
 628                         goto out;
 629                 }
 630
 631                 /*
 632                  * Isolate the page to avoid collapsing an hugepage
 633                  * currently in use by the VM.
 634                  */
 635                 if (!folio_isolate_lru(folio)) {
 636                         folio_unlock(folio);
 637                         result = SCAN_DEL_PAGE_LRU;
 638                         goto out;
 639                 }
 640                 node_stat_mod_folio(folio,
 641                                 NR_ISOLATED_ANON + folio_is_file_lru(folio),
 642                                 folio_nr_pages(folio));
 643                 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
 644                 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
 645
 646                 if (folio_test_large(folio))
 647                         list_add_tail(&folio->lru, compound_pagelist);
 648 next:
 649                 /*
 650                  * If collapse was initiated by khugepaged, check that there is
 651                  * enough young pte to justify collapsing the page
 652                  */
 653                 if (cc->is_khugepaged &&
 654                     (pte_young(pteval) || folio_test_young(folio) ||
 655                      folio_test_referenced(folio) || mmu_notifier_test_young(vma->vm_mm,
 656                                                                      address)))
 657                         referenced++;
 658
 659                 if (pte_write(pteval))
 660                         writable = true;
 661         }
 662
 663         if (unlikely(!writable)) {
 664                 result = SCAN_PAGE_RO;
 665         } else if (unlikely(cc->is_khugepaged && !referenced)) {
 666                 result = SCAN_LACK_REFERENCED_PAGE;
 667         } else {
 668                 result = SCAN_SUCCEED;
 669                 trace_mm_collapse_huge_page_isolate(&folio->page, none_or_zero,
 670                                                     referenced, writable, result);
 671                 return result;
 672         }
 673 out:
 674         release_pte_pages(pte, _pte, compound_pagelist);
 675         trace_mm_collapse_huge_page_isolate(&folio->page, none_or_zero,
 676                                             referenced, writable, result);
 677         return result;
 678 }
 679
 680 static void __collapse_huge_page_copy_succeeded(pte_t *pte,
 681                                                 struct vm_area_struct *vma,
 682                                                 unsigned long address,
 683                                                 spinlock_t *ptl,
 684                                                 struct list_head *compound_pagelist)
 685 {
 686         struct folio *src_folio;
 687         struct page *src_page;
 688         struct page *tmp;
 689         pte_t *_pte;
 690         pte_t pteval;
 691
 692         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 693              _pte++, address += PAGE_SIZE) {
 694                 pteval = ptep_get(_pte);
 695                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
 696                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
 697                         if (is_zero_pfn(pte_pfn(pteval))) {
 698                                 /*
 699                                  * ptl mostly unnecessary.
 700                                  */
 701                                 spin_lock(ptl);
 702                                 ptep_clear(vma->vm_mm, address, _pte);
 703                                 spin_unlock(ptl);
 704                                 ksm_might_unmap_zero_page(vma->vm_mm, pteval);
 705                         }
 706                 } else {
 707                         src_page = pte_page(pteval);
 708                         src_folio = page_folio(src_page);
 709                         if (!folio_test_large(src_folio))
 710                                 release_pte_folio(src_folio);
 711                         /*
 712                          * ptl mostly unnecessary, but preempt has to
 713                          * be disabled to update the per-cpu stats
 714                          * inside folio_remove_rmap_pte().
 715                          */
 716                         spin_lock(ptl);
 717                         ptep_clear(vma->vm_mm, address, _pte);
 718                         folio_remove_rmap_pte(src_folio, src_page, vma);
 719                         spin_unlock(ptl);
 720                         free_page_and_swap_cache(src_page);
 721                 }
 722         }
 723
 724         list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
 725                 list_del(&src_page->lru);
 726                 mod_node_page_state(page_pgdat(src_page),
 727                                     NR_ISOLATED_ANON + page_is_file_lru(src_page),
 728                                     -compound_nr(src_page));
 729                 unlock_page(src_page);
 730                 free_swap_cache(src_page);
 731                 putback_lru_page(src_page);
 732         }
 733 }
 734
 735 static void __collapse_huge_page_copy_failed(pte_t *pte,
 736                                              pmd_t *pmd,
 737                                              pmd_t orig_pmd,
 738                                              struct vm_area_struct *vma,
 739                                              struct list_head *compound_pagelist)
 740 {
 741         spinlock_t *pmd_ptl;
 742
 743         /*
 744          * Re-establish the PMD to point to the original page table
 745          * entry. Restoring PMD needs to be done prior to releasing
 746          * pages. Since pages are still isolated and locked here,
 747          * acquiring anon_vma_lock_write is unnecessary.
 748          */
 749         pmd_ptl = pmd_lock(vma->vm_mm, pmd);
 750         pmd_populate(vma->vm_mm, pmd, pmd_pgtable(orig_pmd));
 751         spin_unlock(pmd_ptl);
 752         /*
 753          * Release both raw and compound pages isolated
 754          * in __collapse_huge_page_isolate.
 755          */
 756         release_pte_pages(pte, pte + HPAGE_PMD_NR, compound_pagelist);
 757 }
 758
 759 /*
 760  * __collapse_huge_page_copy - attempts to copy memory contents from raw
 761  * pages to a hugepage. Cleans up the raw pages if copying succeeds;
 762  * otherwise restores the original page table and releases isolated raw pages.
 763  * Returns SCAN_SUCCEED if copying succeeds, otherwise returns SCAN_COPY_MC.
 764  *
 765  * @pte: starting of the PTEs to copy from
 766  * @page: the new hugepage to copy contents to
 767  * @pmd: pointer to the new hugepage's PMD
 768  * @orig_pmd: the original raw pages' PMD
 769  * @vma: the original raw pages' virtual memory area
 770  * @address: starting address to copy
 771  * @ptl: lock on raw pages' PTEs
 772  * @compound_pagelist: list that stores compound pages
 773  */
 774 static int __collapse_huge_page_copy(pte_t *pte,
 775                                      struct page *page,
 776                                      pmd_t *pmd,
 777                                      pmd_t orig_pmd,
 778                                      struct vm_area_struct *vma,
 779                                      unsigned long address,
 780                                      spinlock_t *ptl,
 781                                      struct list_head *compound_pagelist)
 782 {
 783         struct page *src_page;
 784         pte_t *_pte;
 785         pte_t pteval;
 786         unsigned long _address;
 787         int result = SCAN_SUCCEED;
 788
 789         /*
 790          * Copying pages' contents is subject to memory poison at any iteration.
 791          */
 792         for (_pte = pte, _address = address; _pte < pte + HPAGE_PMD_NR;
 793              _pte++, page++, _address += PAGE_SIZE) {
 794                 pteval = ptep_get(_pte);
 795                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
 796                         clear_user_highpage(page, _address);
 797                         continue;
 798                 }
 799                 src_page = pte_page(pteval);
 800                 if (copy_mc_user_highpage(page, src_page, _address, vma) > 0) {
 801                         result = SCAN_COPY_MC;
 802                         break;
 803                 }
 804         }
 805
 806         if (likely(result == SCAN_SUCCEED))
 807                 __collapse_huge_page_copy_succeeded(pte, vma, address, ptl,
 808                                                     compound_pagelist);
 809         else
 810                 __collapse_huge_page_copy_failed(pte, pmd, orig_pmd, vma,
 811                                                  compound_pagelist);
 812
 813         return result;
 814 }
 815
 816 static void khugepaged_alloc_sleep(void)
 817 {
 818         DEFINE_WAIT(wait);
 819
 820         add_wait_queue(&khugepaged_wait, &wait);
 821         __set_current_state(TASK_INTERRUPTIBLE|TASK_FREEZABLE);
 822         schedule_timeout(msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
 823         remove_wait_queue(&khugepaged_wait, &wait);
 824 }
 825
 826 struct collapse_control khugepaged_collapse_control = {
 827         .is_khugepaged = true,
 828 };
 829
 830 static bool hpage_collapse_scan_abort(int nid, struct collapse_control *cc)
 831 {
 832         int i;
 833
 834         /*
 835          * If node_reclaim_mode is disabled, then no extra effort is made to
 836          * allocate memory locally.
 837          */
 838         if (!node_reclaim_enabled())
 839                 return false;
 840
 841         /* If there is a count for this node already, it must be acceptable */
 842         if (cc->node_load[nid])
 843                 return false;
 844
 845         for (i = 0; i < MAX_NUMNODES; i++) {
 846                 if (!cc->node_load[i])
 847                         continue;
 848                 if (node_distance(nid, i) > node_reclaim_distance)
 849                         return true;
 850         }
 851         return false;
 852 }
 853
 854 #define khugepaged_defrag()                                     \
 855         (transparent_hugepage_flags &                           \
 856          (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG))
 857
 858 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
 859 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
 860 {
 861         return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
 862 }
 863
 864 #ifdef CONFIG_NUMA
 865 static int hpage_collapse_find_target_node(struct collapse_control *cc)
 866 {
 867         int nid, target_node = 0, max_value = 0;
 868
 869         /* find first node with max normal pages hit */
 870         for (nid = 0; nid < MAX_NUMNODES; nid++)
 871                 if (cc->node_load[nid] > max_value) {
 872                         max_value = cc->node_load[nid];
 873                         target_node = nid;
 874                 }
 875
 876         for_each_online_node(nid) {
 877                 if (max_value == cc->node_load[nid])
 878                         node_set(nid, cc->alloc_nmask);
 879         }
 880
 881         return target_node;
 882 }
 883 #else
 884 static int hpage_collapse_find_target_node(struct collapse_control *cc)
 885 {
 886         return 0;
 887 }
 888 #endif
 889
 890 static bool hpage_collapse_alloc_folio(struct folio **folio, gfp_t gfp, int node,
 891                                       nodemask_t *nmask)
 892 {
 893         *folio = __folio_alloc(gfp, HPAGE_PMD_ORDER, node, nmask);
 894
 895         if (unlikely(!*folio)) {
 896                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 897                 return false;
 898         }
 899
 900         count_vm_event(THP_COLLAPSE_ALLOC);
 901         return true;
 902 }
 903
 904 /*
 905  * If mmap_lock temporarily dropped, revalidate vma
 906  * before taking mmap_lock.
 907  * Returns enum scan_result value.
 908  */
 909
 910 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
 911                                    bool expect_anon,
 912                                    struct vm_area_struct **vmap,
 913                                    struct collapse_control *cc)
 914 {
 915         struct vm_area_struct *vma;
 916
 917         if (unlikely(hpage_collapse_test_exit(mm)))
 918                 return SCAN_ANY_PROCESS;
 919
 920         *vmap = vma = find_vma(mm, address);
 921         if (!vma)
 922                 return SCAN_VMA_NULL;
 923
 924         if (!thp_vma_suitable_order(vma, address, PMD_ORDER))
 925                 return SCAN_ADDRESS_RANGE;
 926         if (!thp_vma_allowable_order(vma, vma->vm_flags, false, false,
 927                                      cc->is_khugepaged, PMD_ORDER))
 928                 return SCAN_VMA_CHECK;
 929         /*
 930          * Anon VMA expected, the address may be unmapped then
 931          * remapped to file after khugepaged reaquired the mmap_lock.
 932          *
 933          * thp_vma_allowable_order may return true for qualified file
 934          * vmas.
 935          */
 936         if (expect_anon && (!(*vmap)->anon_vma || !vma_is_anonymous(*vmap)))
 937                 return SCAN_PAGE_ANON;
 938         return SCAN_SUCCEED;
 939 }
 940
 941 static int find_pmd_or_thp_or_none(struct mm_struct *mm,
 942                                    unsigned long address,
 943                                    pmd_t **pmd)
 944 {
 945         pmd_t pmde;
 946
 947         *pmd = mm_find_pmd(mm, address);
 948         if (!*pmd)
 949                 return SCAN_PMD_NULL;
 950
 951         pmde = pmdp_get_lockless(*pmd);
 952         if (pmd_none(pmde))
 953                 return SCAN_PMD_NONE;
 954         if (!pmd_present(pmde))
 955                 return SCAN_PMD_NULL;
 956         if (pmd_trans_huge(pmde))
 957                 return SCAN_PMD_MAPPED;
 958         if (pmd_devmap(pmde))
 959                 return SCAN_PMD_NULL;
 960         if (pmd_bad(pmde))
 961                 return SCAN_PMD_NULL;
 962         return SCAN_SUCCEED;
 963 }
 964
 965 static int check_pmd_still_valid(struct mm_struct *mm,
 966                                  unsigned long address,
 967                                  pmd_t *pmd)
 968 {
 969         pmd_t *new_pmd;
 970         int result = find_pmd_or_thp_or_none(mm, address, &new_pmd);
 971
 972         if (result != SCAN_SUCCEED)
 973                 return result;
 974         if (new_pmd != pmd)
 975                 return SCAN_FAIL;
 976         return SCAN_SUCCEED;
 977 }
 978
 979 /*
 980  * Bring missing pages in from swap, to complete THP collapse.
 981  * Only done if hpage_collapse_scan_pmd believes it is worthwhile.
 982  *
 983  * Called and returns without pte mapped or spinlocks held.
 984  * Returns result: if not SCAN_SUCCEED, mmap_lock has been released.
 985  */
 986 static int __collapse_huge_page_swapin(struct mm_struct *mm,
 987                                        struct vm_area_struct *vma,
 988                                        unsigned long haddr, pmd_t *pmd,
 989                                        int referenced)
 990 {
 991         int swapped_in = 0;
 992         vm_fault_t ret = 0;
 993         unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE);
 994         int result;
 995         pte_t *pte = NULL;
 996         spinlock_t *ptl;
 997
 998         for (address = haddr; address < end; address += PAGE_SIZE) {
 999                 struct vm_fault vmf = {
1000                         .vma = vma,
1001                         .address = address,
1002                         .pgoff = linear_page_index(vma, address),
1003                         .flags = FAULT_FLAG_ALLOW_RETRY,
1004                         .pmd = pmd,
1005                 };
1006
1007                 if (!pte++) {
1008                         pte = pte_offset_map_nolock(mm, pmd, address, &ptl);
1009                         if (!pte) {
1010                                 mmap_read_unlock(mm);
1011                                 result = SCAN_PMD_NULL;
1012                                 goto out;
1013                         }
1014                 }
1015
1016                 vmf.orig_pte = ptep_get_lockless(pte);
1017                 if (!is_swap_pte(vmf.orig_pte))
1018                         continue;
1019
1020                 vmf.pte = pte;
1021                 vmf.ptl = ptl;
1022                 ret = do_swap_page(&vmf);
1023                 /* Which unmaps pte (after perhaps re-checking the entry) */
1024                 pte = NULL;
1025
1026                 /*
1027                  * do_swap_page returns VM_FAULT_RETRY with released mmap_lock.
1028                  * Note we treat VM_FAULT_RETRY as VM_FAULT_ERROR here because
1029                  * we do not retry here and swap entry will remain in pagetable
1030                  * resulting in later failure.
1031                  */
1032                 if (ret & VM_FAULT_RETRY) {
1033                         /* Likely, but not guaranteed, that page lock failed */
1034                         result = SCAN_PAGE_LOCK;
1035                         goto out;
1036                 }
1037                 if (ret & VM_FAULT_ERROR) {
1038                         mmap_read_unlock(mm);
1039                         result = SCAN_FAIL;
1040                         goto out;
1041                 }
1042                 swapped_in++;
1043         }
1044
1045         if (pte)
1046                 pte_unmap(pte);
1047
1048         /* Drain LRU cache to remove extra pin on the swapped in pages */
1049         if (swapped_in)
1050                 lru_add_drain();
1051
1052         result = SCAN_SUCCEED;
1053 out:
1054         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, result);
1055         return result;
1056 }
1057
1058 static int alloc_charge_hpage(struct page **hpage, struct mm_struct *mm,
1059                               struct collapse_control *cc)
1060 {
1061         gfp_t gfp = (cc->is_khugepaged ? alloc_hugepage_khugepaged_gfpmask() :
1062                      GFP_TRANSHUGE);
1063         int node = hpage_collapse_find_target_node(cc);
1064         struct folio *folio;
1065
1066         if (!hpage_collapse_alloc_folio(&folio, gfp, node, &cc->alloc_nmask)) {
1067                 *hpage = NULL;
1068                 return SCAN_ALLOC_HUGE_PAGE_FAIL;
1069         }
1070
1071         if (unlikely(mem_cgroup_charge(folio, mm, gfp))) {
1072                 folio_put(folio);
1073                 *hpage = NULL;
1074                 return SCAN_CGROUP_CHARGE_FAIL;
1075         }
1076
1077         count_memcg_folio_events(folio, THP_COLLAPSE_ALLOC, 1);
1078
1079         *hpage = folio_page(folio, 0);
1080         return SCAN_SUCCEED;
1081 }
1082
1083 static int collapse_huge_page(struct mm_struct *mm, unsigned long address,
1084                               int referenced, int unmapped,
1085                               struct collapse_control *cc)
1086 {
1087         LIST_HEAD(compound_pagelist);
1088         pmd_t *pmd, _pmd;
1089         pte_t *pte;
1090         pgtable_t pgtable;
1091         struct folio *folio;
1092         struct page *hpage;
1093         spinlock_t *pmd_ptl, *pte_ptl;
1094         int result = SCAN_FAIL;
1095         struct vm_area_struct *vma;
1096         struct mmu_notifier_range range;
1097
1098         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1099
1100         /*
1101          * Before allocating the hugepage, release the mmap_lock read lock.
1102          * The allocation can take potentially a long time if it involves
1103          * sync compaction, and we do not need to hold the mmap_lock during
1104          * that. We will recheck the vma after taking it again in write mode.
1105          */
1106         mmap_read_unlock(mm);
1107
1108         result = alloc_charge_hpage(&hpage, mm, cc);
1109         if (result != SCAN_SUCCEED)
1110                 goto out_nolock;
1111
1112         mmap_read_lock(mm);
1113         result = hugepage_vma_revalidate(mm, address, true, &vma, cc);
1114         if (result != SCAN_SUCCEED) {
1115                 mmap_read_unlock(mm);
1116                 goto out_nolock;
1117         }
1118
1119         result = find_pmd_or_thp_or_none(mm, address, &pmd);
1120         if (result != SCAN_SUCCEED) {
1121                 mmap_read_unlock(mm);
1122                 goto out_nolock;
1123         }
1124
1125         if (unmapped) {
1126                 /*
1127                  * __collapse_huge_page_swapin will return with mmap_lock
1128                  * released when it fails. So we jump out_nolock directly in
1129                  * that case.  Continuing to collapse causes inconsistency.
1130                  */
1131                 result = __collapse_huge_page_swapin(mm, vma, address, pmd,
1132                                                      referenced);
1133                 if (result != SCAN_SUCCEED)
1134                         goto out_nolock;
1135         }
1136
1137         mmap_read_unlock(mm);
1138         /*
1139          * Prevent all access to pagetables with the exception of
1140          * gup_fast later handled by the ptep_clear_flush and the VM
1141          * handled by the anon_vma lock + PG_lock.
1142          *
1143          * UFFDIO_MOVE is prevented to race as well thanks to the
1144          * mmap_lock.
1145          */
1146         mmap_write_lock(mm);
1147         result = hugepage_vma_revalidate(mm, address, true, &vma, cc);
1148         if (result != SCAN_SUCCEED)
1149                 goto out_up_write;
1150         /* check if the pmd is still valid */
1151         result = check_pmd_still_valid(mm, address, pmd);
1152         if (result != SCAN_SUCCEED)
1153                 goto out_up_write;
1154
1155         vma_start_write(vma);
1156         anon_vma_lock_write(vma->anon_vma);
1157
1158         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, address,
1159                                 address + HPAGE_PMD_SIZE);
1160         mmu_notifier_invalidate_range_start(&range);
1161
1162         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1163         /*
1164          * This removes any huge TLB entry from the CPU so we won't allow
1165          * huge and small TLB entries for the same virtual address to
1166          * avoid the risk of CPU bugs in that area.
1167          *
1168          * Parallel fast GUP is fine since fast GUP will back off when
1169          * it detects PMD is changed.
1170          */
1171         _pmd = pmdp_collapse_flush(vma, address, pmd);
1172         spin_unlock(pmd_ptl);
1173         mmu_notifier_invalidate_range_end(&range);
1174         tlb_remove_table_sync_one();
1175
1176         pte = pte_offset_map_lock(mm, &_pmd, address, &pte_ptl);
1177         if (pte) {
1178                 result = __collapse_huge_page_isolate(vma, address, pte, cc,
1179                                                       &compound_pagelist);
1180                 spin_unlock(pte_ptl);
1181         } else {
1182                 result = SCAN_PMD_NULL;
1183         }
1184
1185         if (unlikely(result != SCAN_SUCCEED)) {
1186                 if (pte)
1187                         pte_unmap(pte);
1188                 spin_lock(pmd_ptl);
1189                 BUG_ON(!pmd_none(*pmd));
1190                 /*
1191                  * We can only use set_pmd_at when establishing
1192                  * hugepmds and never for establishing regular pmds that
1193                  * points to regular pagetables. Use pmd_populate for that
1194                  */
1195                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1196                 spin_unlock(pmd_ptl);
1197                 anon_vma_unlock_write(vma->anon_vma);
1198                 goto out_up_write;
1199         }
1200
1201         /*
1202          * All pages are isolated and locked so anon_vma rmap
1203          * can't run anymore.
1204          */
1205         anon_vma_unlock_write(vma->anon_vma);
1206
1207         result = __collapse_huge_page_copy(pte, hpage, pmd, _pmd,
1208                                            vma, address, pte_ptl,
1209                                            &compound_pagelist);
1210         pte_unmap(pte);
1211         if (unlikely(result != SCAN_SUCCEED))
1212                 goto out_up_write;
1213
1214         folio = page_folio(hpage);
1215         /*
1216          * The smp_wmb() inside __folio_mark_uptodate() ensures the
1217          * copy_huge_page writes become visible before the set_pmd_at()
1218          * write.
1219          */
1220         __folio_mark_uptodate(folio);
1221         pgtable = pmd_pgtable(_pmd);
1222
1223         _pmd = mk_huge_pmd(hpage, vma->vm_page_prot);
1224         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1225
1226         spin_lock(pmd_ptl);
1227         BUG_ON(!pmd_none(*pmd));
1228         folio_add_new_anon_rmap(folio, vma, address);
1229         folio_add_lru_vma(folio, vma);
1230         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1231         set_pmd_at(mm, address, pmd, _pmd);
1232         update_mmu_cache_pmd(vma, address, pmd);
1233         spin_unlock(pmd_ptl);
1234
1235         hpage = NULL;
1236
1237         result = SCAN_SUCCEED;
1238 out_up_write:
1239         mmap_write_unlock(mm);
1240 out_nolock:
1241         if (hpage)
1242                 put_page(hpage);
1243         trace_mm_collapse_huge_page(mm, result == SCAN_SUCCEED, result);
1244         return result;
1245 }
1246
1247 static int hpage_collapse_scan_pmd(struct mm_struct *mm,
1248                                    struct vm_area_struct *vma,
1249                                    unsigned long address, bool *mmap_locked,
1250                                    struct collapse_control *cc)
1251 {
1252         pmd_t *pmd;
1253         pte_t *pte, *_pte;
1254         int result = SCAN_FAIL, referenced = 0;
1255         int none_or_zero = 0, shared = 0;
1256         struct page *page = NULL;
1257         struct folio *folio = NULL;
1258         unsigned long _address;
1259         spinlock_t *ptl;
1260         int node = NUMA_NO_NODE, unmapped = 0;
1261         bool writable = false;
1262
1263         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1264
1265         result = find_pmd_or_thp_or_none(mm, address, &pmd);
1266         if (result != SCAN_SUCCEED)
1267                 goto out;
1268
1269         memset(cc->node_load, 0, sizeof(cc->node_load));
1270         nodes_clear(cc->alloc_nmask);
1271         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1272         if (!pte) {
1273                 result = SCAN_PMD_NULL;
1274                 goto out;
1275         }
1276
1277         for (_address = address, _pte = pte; _pte < pte + HPAGE_PMD_NR;
1278              _pte++, _address += PAGE_SIZE) {
1279                 pte_t pteval = ptep_get(_pte);
1280                 if (is_swap_pte(pteval)) {
1281                         ++unmapped;
1282                         if (!cc->is_khugepaged ||
1283                             unmapped <= khugepaged_max_ptes_swap) {
1284                                 /*
1285                                  * Always be strict with uffd-wp
1286                                  * enabled swap entries.  Please see
1287                                  * comment below for pte_uffd_wp().
1288                                  */
1289                                 if (pte_swp_uffd_wp_any(pteval)) {
1290                                         result = SCAN_PTE_UFFD_WP;
1291                                         goto out_unmap;
1292                                 }
1293                                 continue;
1294                         } else {
1295                                 result = SCAN_EXCEED_SWAP_PTE;
1296                                 count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
1297                                 goto out_unmap;
1298                         }
1299                 }
1300                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1301                         ++none_or_zero;
1302                         if (!userfaultfd_armed(vma) &&
1303                             (!cc->is_khugepaged ||
1304                              none_or_zero <= khugepaged_max_ptes_none)) {
1305                                 continue;
1306                         } else {
1307                                 result = SCAN_EXCEED_NONE_PTE;
1308                                 count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
1309                                 goto out_unmap;
1310                         }
1311                 }
1312                 if (pte_uffd_wp(pteval)) {
1313                         /*
1314                          * Don't collapse the page if any of the small
1315                          * PTEs are armed with uffd write protection.
1316                          * Here we can also mark the new huge pmd as
1317                          * write protected if any of the small ones is
1318                          * marked but that could bring unknown
1319                          * userfault messages that falls outside of
1320                          * the registered range.  So, just be simple.
1321                          */
1322                         result = SCAN_PTE_UFFD_WP;
1323                         goto out_unmap;
1324                 }
1325                 if (pte_write(pteval))
1326                         writable = true;
1327
1328                 page = vm_normal_page(vma, _address, pteval);
1329                 if (unlikely(!page) || unlikely(is_zone_device_page(page))) {
1330                         result = SCAN_PAGE_NULL;
1331                         goto out_unmap;
1332                 }
1333
1334                 if (page_mapcount(page) > 1) {
1335                         ++shared;
1336                         if (cc->is_khugepaged &&
1337                             shared > khugepaged_max_ptes_shared) {
1338                                 result = SCAN_EXCEED_SHARED_PTE;
1339                                 count_vm_event(THP_SCAN_EXCEED_SHARED_PTE);
1340                                 goto out_unmap;
1341                         }
1342                 }
1343
1344                 folio = page_folio(page);
1345                 /*
1346                  * Record which node the original page is from and save this
1347                  * information to cc->node_load[].
1348                  * Khugepaged will allocate hugepage from the node has the max
1349                  * hit record.
1350                  */
1351                 node = folio_nid(folio);
1352                 if (hpage_collapse_scan_abort(node, cc)) {
1353                         result = SCAN_SCAN_ABORT;
1354                         goto out_unmap;
1355                 }
1356                 cc->node_load[node]++;
1357                 if (!folio_test_lru(folio)) {
1358                         result = SCAN_PAGE_LRU;
1359                         goto out_unmap;
1360                 }
1361                 if (folio_test_locked(folio)) {
1362                         result = SCAN_PAGE_LOCK;
1363                         goto out_unmap;
1364                 }
1365                 if (!folio_test_anon(folio)) {
1366                         result = SCAN_PAGE_ANON;
1367                         goto out_unmap;
1368                 }
1369
1370                 /*
1371                  * Check if the page has any GUP (or other external) pins.
1372                  *
1373                  * Here the check may be racy:
1374                  * it may see total_mapcount > refcount in some cases?
1375                  * But such case is ephemeral we could always retry collapse
1376                  * later.  However it may report false positive if the page
1377                  * has excessive GUP pins (i.e. 512).  Anyway the same check
1378                  * will be done again later the risk seems low.
1379                  */
1380                 if (!is_refcount_suitable(folio)) {
1381                         result = SCAN_PAGE_COUNT;
1382                         goto out_unmap;
1383                 }
1384
1385                 /*
1386                  * If collapse was initiated by khugepaged, check that there is
1387                  * enough young pte to justify collapsing the page
1388                  */
1389                 if (cc->is_khugepaged &&
1390                     (pte_young(pteval) || folio_test_young(folio) ||
1391                      folio_test_referenced(folio) || mmu_notifier_test_young(vma->vm_mm,
1392                                                                      address)))
1393                         referenced++;
1394         }
1395         if (!writable) {
1396                 result = SCAN_PAGE_RO;
1397         } else if (cc->is_khugepaged &&
1398                    (!referenced ||
1399                     (unmapped && referenced < HPAGE_PMD_NR / 2))) {
1400                 result = SCAN_LACK_REFERENCED_PAGE;
1401         } else {
1402                 result = SCAN_SUCCEED;
1403         }
1404 out_unmap:
1405         pte_unmap_unlock(pte, ptl);
1406         if (result == SCAN_SUCCEED) {
1407                 result = collapse_huge_page(mm, address, referenced,
1408                                             unmapped, cc);
1409                 /* collapse_huge_page will return with the mmap_lock released */
1410                 *mmap_locked = false;
1411         }
1412 out:
1413         trace_mm_khugepaged_scan_pmd(mm, &folio->page, writable, referenced,
1414                                      none_or_zero, result, unmapped);
1415         return result;
1416 }
1417
1418 static void collect_mm_slot(struct khugepaged_mm_slot *mm_slot)
1419 {
1420         struct mm_slot *slot = &mm_slot->slot;
1421         struct mm_struct *mm = slot->mm;
1422
1423         lockdep_assert_held(&khugepaged_mm_lock);
1424
1425         if (hpage_collapse_test_exit(mm)) {
1426                 /* free mm_slot */
1427                 hash_del(&slot->hash);
1428                 list_del(&slot->mm_node);
1429
1430                 /*
1431                  * Not strictly needed because the mm exited already.
1432                  *
1433                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1434                  */
1435
1436                 /* khugepaged_mm_lock actually not necessary for the below */
1437                 mm_slot_free(mm_slot_cache, mm_slot);
1438                 mmdrop(mm);
1439         }
1440 }
1441
1442 #ifdef CONFIG_SHMEM
1443 /* hpage must be locked, and mmap_lock must be held */
1444 static int set_huge_pmd(struct vm_area_struct *vma, unsigned long addr,
1445                         pmd_t *pmdp, struct page *hpage)
1446 {
1447         struct vm_fault vmf = {
1448                 .vma = vma,
1449                 .address = addr,
1450                 .flags = 0,
1451                 .pmd = pmdp,
1452         };
1453
1454         VM_BUG_ON(!PageTransHuge(hpage));
1455         mmap_assert_locked(vma->vm_mm);
1456
1457         if (do_set_pmd(&vmf, hpage))
1458                 return SCAN_FAIL;
1459
1460         get_page(hpage);
1461         return SCAN_SUCCEED;
1462 }
1463
1464 /**
1465  * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at
1466  * address haddr.
1467  *
1468  * @mm: process address space where collapse happens
1469  * @addr: THP collapse address
1470  * @install_pmd: If a huge PMD should be installed
1471  *
1472  * This function checks whether all the PTEs in the PMD are pointing to the
1473  * right THP. If so, retract the page table so the THP can refault in with
1474  * as pmd-mapped. Possibly install a huge PMD mapping the THP.
1475  */
1476 int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
1477                             bool install_pmd)
1478 {
1479         struct mmu_notifier_range range;
1480         bool notified = false;
1481         unsigned long haddr = addr & HPAGE_PMD_MASK;
1482         struct vm_area_struct *vma = vma_lookup(mm, haddr);
1483         struct folio *folio;
1484         pte_t *start_pte, *pte;
1485         pmd_t *pmd, pgt_pmd;
1486         spinlock_t *pml = NULL, *ptl;
1487         int nr_ptes = 0, result = SCAN_FAIL;
1488         int i;
1489
1490         mmap_assert_locked(mm);
1491
1492         /* First check VMA found, in case page tables are being torn down */
1493         if (!vma || !vma->vm_file ||
1494             !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE))
1495                 return SCAN_VMA_CHECK;
1496
1497         /* Fast check before locking page if already PMD-mapped */
1498         result = find_pmd_or_thp_or_none(mm, haddr, &pmd);
1499         if (result == SCAN_PMD_MAPPED)
1500                 return result;
1501
1502         /*
1503          * If we are here, we've succeeded in replacing all the native pages
1504          * in the page cache with a single hugepage. If a mm were to fault-in
1505          * this memory (mapped by a suitably aligned VMA), we'd get the hugepage
1506          * and map it by a PMD, regardless of sysfs THP settings. As such, let's
1507          * analogously elide sysfs THP settings here.
1508          */
1509         if (!thp_vma_allowable_order(vma, vma->vm_flags, false, false, false,
1510                                      PMD_ORDER))
1511                 return SCAN_VMA_CHECK;
1512
1513         /* Keep pmd pgtable for uffd-wp; see comment in retract_page_tables() */
1514         if (userfaultfd_wp(vma))
1515                 return SCAN_PTE_UFFD_WP;
1516
1517         folio = filemap_lock_folio(vma->vm_file->f_mapping,
1518                                linear_page_index(vma, haddr));
1519         if (IS_ERR(folio))
1520                 return SCAN_PAGE_NULL;
1521
1522         if (folio_order(folio) != HPAGE_PMD_ORDER) {
1523                 result = SCAN_PAGE_COMPOUND;
1524                 goto drop_folio;
1525         }
1526
1527         result = find_pmd_or_thp_or_none(mm, haddr, &pmd);
1528         switch (result) {
1529         case SCAN_SUCCEED:
1530                 break;
1531         case SCAN_PMD_NONE:
1532                 /*
1533                  * All pte entries have been removed and pmd cleared.
1534                  * Skip all the pte checks and just update the pmd mapping.
1535                  */
1536                 goto maybe_install_pmd;
1537         default:
1538                 goto drop_folio;
1539         }
1540
1541         result = SCAN_FAIL;
1542         start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1543         if (!start_pte)         /* mmap_lock + page lock should prevent this */
1544                 goto drop_folio;
1545
1546         /* step 1: check all mapped PTEs are to the right huge page */
1547         for (i = 0, addr = haddr, pte = start_pte;
1548              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1549                 struct page *page;
1550                 pte_t ptent = ptep_get(pte);
1551
1552                 /* empty pte, skip */
1553                 if (pte_none(ptent))
1554                         continue;
1555
1556                 /* page swapped out, abort */
1557                 if (!pte_present(ptent)) {
1558                         result = SCAN_PTE_NON_PRESENT;
1559                         goto abort;
1560                 }
1561
1562                 page = vm_normal_page(vma, addr, ptent);
1563                 if (WARN_ON_ONCE(page && is_zone_device_page(page)))
1564                         page = NULL;
1565                 /*
1566                  * Note that uprobe, debugger, or MAP_PRIVATE may change the
1567                  * page table, but the new page will not be a subpage of hpage.
1568                  */
1569                 if (folio_page(folio, i) != page)
1570                         goto abort;
1571         }
1572
1573         pte_unmap_unlock(start_pte, ptl);
1574         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
1575                                 haddr, haddr + HPAGE_PMD_SIZE);
1576         mmu_notifier_invalidate_range_start(&range);
1577         notified = true;
1578
1579         /*
1580          * pmd_lock covers a wider range than ptl, and (if split from mm's
1581          * page_table_lock) ptl nests inside pml. The less time we hold pml,
1582          * the better; but userfaultfd's mfill_atomic_pte() on a private VMA
1583          * inserts a valid as-if-COWed PTE without even looking up page cache.
1584          * So page lock of folio does not protect from it, so we must not drop
1585          * ptl before pgt_pmd is removed, so uffd private needs pml taken now.
1586          */
1587         if (userfaultfd_armed(vma) && !(vma->vm_flags & VM_SHARED))
1588                 pml = pmd_lock(mm, pmd);
1589
1590         start_pte = pte_offset_map_nolock(mm, pmd, haddr, &ptl);
1591         if (!start_pte)         /* mmap_lock + page lock should prevent this */
1592                 goto abort;
1593         if (!pml)
1594                 spin_lock(ptl);
1595         else if (ptl != pml)
1596                 spin_lock_nested(ptl, SINGLE_DEPTH_NESTING);
1597
1598         /* step 2: clear page table and adjust rmap */
1599         for (i = 0, addr = haddr, pte = start_pte;
1600              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1601                 struct page *page;
1602                 pte_t ptent = ptep_get(pte);
1603
1604                 if (pte_none(ptent))
1605                         continue;
1606                 /*
1607                  * We dropped ptl after the first scan, to do the mmu_notifier:
1608                  * page lock stops more PTEs of the folio being faulted in, but
1609                  * does not stop write faults COWing anon copies from existing
1610                  * PTEs; and does not stop those being swapped out or migrated.
1611                  */
1612                 if (!pte_present(ptent)) {
1613                         result = SCAN_PTE_NON_PRESENT;
1614                         goto abort;
1615                 }
1616                 page = vm_normal_page(vma, addr, ptent);
1617                 if (folio_page(folio, i) != page)
1618                         goto abort;
1619
1620                 /*
1621                  * Must clear entry, or a racing truncate may re-remove it.
1622                  * TLB flush can be left until pmdp_collapse_flush() does it.
1623                  * PTE dirty? Shmem page is already dirty; file is read-only.
1624                  */
1625                 ptep_clear(mm, addr, pte);
1626                 folio_remove_rmap_pte(folio, page, vma);
1627                 nr_ptes++;
1628         }
1629
1630         pte_unmap(start_pte);
1631         if (!pml)
1632                 spin_unlock(ptl);
1633
1634         /* step 3: set proper refcount and mm_counters. */
1635         if (nr_ptes) {
1636                 folio_ref_sub(folio, nr_ptes);
1637                 add_mm_counter(mm, mm_counter_file(&folio->page), -nr_ptes);
1638         }
1639
1640         /* step 4: remove empty page table */
1641         if (!pml) {
1642                 pml = pmd_lock(mm, pmd);
1643                 if (ptl != pml)
1644                         spin_lock_nested(ptl, SINGLE_DEPTH_NESTING);
1645         }
1646         pgt_pmd = pmdp_collapse_flush(vma, haddr, pmd);
1647         pmdp_get_lockless_sync();
1648         if (ptl != pml)
1649                 spin_unlock(ptl);
1650         spin_unlock(pml);
1651
1652         mmu_notifier_invalidate_range_end(&range);
1653
1654         mm_dec_nr_ptes(mm);
1655         page_table_check_pte_clear_range(mm, haddr, pgt_pmd);
1656         pte_free_defer(mm, pmd_pgtable(pgt_pmd));
1657
1658 maybe_install_pmd:
1659         /* step 5: install pmd entry */
1660         result = install_pmd
1661                         ? set_huge_pmd(vma, haddr, pmd, &folio->page)
1662                         : SCAN_SUCCEED;
1663         goto drop_folio;
1664 abort:
1665         if (nr_ptes) {
1666                 flush_tlb_mm(mm);
1667                 folio_ref_sub(folio, nr_ptes);
1668                 add_mm_counter(mm, mm_counter_file(&folio->page), -nr_ptes);
1669         }
1670         if (start_pte)
1671                 pte_unmap_unlock(start_pte, ptl);
1672         if (pml && pml != ptl)
1673                 spin_unlock(pml);
1674         if (notified)
1675                 mmu_notifier_invalidate_range_end(&range);
1676 drop_folio:
1677         folio_unlock(folio);
1678         folio_put(folio);
1679         return result;
1680 }
1681
1682 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1683 {
1684         struct vm_area_struct *vma;
1685
1686         i_mmap_lock_read(mapping);
1687         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1688                 struct mmu_notifier_range range;
1689                 struct mm_struct *mm;
1690                 unsigned long addr;
1691                 pmd_t *pmd, pgt_pmd;
1692                 spinlock_t *pml;
1693                 spinlock_t *ptl;
1694                 bool skipped_uffd = false;
1695
1696                 /*
1697                  * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1698                  * got written to. These VMAs are likely not worth removing
1699                  * page tables from, as PMD-mapping is likely to be split later.
1700                  */
1701                 if (READ_ONCE(vma->anon_vma))
1702                         continue;
1703
1704                 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1705                 if (addr & ~HPAGE_PMD_MASK ||
1706                     vma->vm_end < addr + HPAGE_PMD_SIZE)
1707                         continue;
1708
1709                 mm = vma->vm_mm;
1710                 if (find_pmd_or_thp_or_none(mm, addr, &pmd) != SCAN_SUCCEED)
1711                         continue;
1712
1713                 if (hpage_collapse_test_exit(mm))
1714                         continue;
1715                 /*
1716                  * When a vma is registered with uffd-wp, we cannot recycle
1717                  * the page table because there may be pte markers installed.
1718                  * Other vmas can still have the same file mapped hugely, but
1719                  * skip this one: it will always be mapped in small page size
1720                  * for uffd-wp registered ranges.
1721                  */
1722                 if (userfaultfd_wp(vma))
1723                         continue;
1724
1725                 /* PTEs were notified when unmapped; but now for the PMD? */
1726                 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
1727                                         addr, addr + HPAGE_PMD_SIZE);
1728                 mmu_notifier_invalidate_range_start(&range);
1729
1730                 pml = pmd_lock(mm, pmd);
1731                 ptl = pte_lockptr(mm, pmd);
1732                 if (ptl != pml)
1733                         spin_lock_nested(ptl, SINGLE_DEPTH_NESTING);
1734
1735                 /*
1736                  * Huge page lock is still held, so normally the page table
1737                  * must remain empty; and we have already skipped anon_vma
1738                  * and userfaultfd_wp() vmas.  But since the mmap_lock is not
1739                  * held, it is still possible for a racing userfaultfd_ioctl()
1740                  * to have inserted ptes or markers.  Now that we hold ptlock,
1741                  * repeating the anon_vma check protects from one category,
1742                  * and repeating the userfaultfd_wp() check from another.
1743                  */
1744                 if (unlikely(vma->anon_vma || userfaultfd_wp(vma))) {
1745                         skipped_uffd = true;
1746                 } else {
1747                         pgt_pmd = pmdp_collapse_flush(vma, addr, pmd);
1748                         pmdp_get_lockless_sync();
1749                 }
1750
1751                 if (ptl != pml)
1752                         spin_unlock(ptl);
1753                 spin_unlock(pml);
1754
1755                 mmu_notifier_invalidate_range_end(&range);
1756
1757                 if (!skipped_uffd) {
1758                         mm_dec_nr_ptes(mm);
1759                         page_table_check_pte_clear_range(mm, addr, pgt_pmd);
1760                         pte_free_defer(mm, pmd_pgtable(pgt_pmd));
1761                 }
1762         }
1763         i_mmap_unlock_read(mapping);
1764 }
1765
1766 /**
1767  * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1768  *
1769  * @mm: process address space where collapse happens
1770  * @addr: virtual collapse start address
1771  * @file: file that collapse on
1772  * @start: collapse start address
1773  * @cc: collapse context and scratchpad
1774  *
1775  * Basic scheme is simple, details are more complex:
1776  *  - allocate and lock a new huge page;
1777  *  - scan page cache, locking old pages
1778  *    + swap/gup in pages if necessary;
1779  *  - copy data to new page
1780  *  - handle shmem holes
1781  *    + re-validate that holes weren't filled by someone else
1782  *    + check for userfaultfd
1783  *  - finalize updates to the page cache;
1784  *  - if replacing succeeds:
1785  *    + unlock huge page;
1786  *    + free old pages;
1787  *  - if replacing failed;
1788  *    + unlock old pages
1789  *    + unlock and free huge page;
1790  */
1791 static int collapse_file(struct mm_struct *mm, unsigned long addr,
1792                          struct file *file, pgoff_t start,
1793                          struct collapse_control *cc)
1794 {
1795         struct address_space *mapping = file->f_mapping;
1796         struct page *hpage;
1797         struct page *page;
1798         struct page *tmp;
1799         struct folio *folio;
1800         pgoff_t index = 0, end = start + HPAGE_PMD_NR;
1801         LIST_HEAD(pagelist);
1802         XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1803         int nr_none = 0, result = SCAN_SUCCEED;
1804         bool is_shmem = shmem_file(file);
1805         int nr = 0;
1806
1807         VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1808         VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1809
1810         result = alloc_charge_hpage(&hpage, mm, cc);
1811         if (result != SCAN_SUCCEED)
1812                 goto out;
1813
1814         __SetPageLocked(hpage);
1815         if (is_shmem)
1816                 __SetPageSwapBacked(hpage);
1817         hpage->index = start;
1818         hpage->mapping = mapping;
1819
1820         /*
1821          * Ensure we have slots for all the pages in the range.  This is
1822          * almost certainly a no-op because most of the pages must be present
1823          */
1824         do {
1825                 xas_lock_irq(&xas);
1826                 xas_create_range(&xas);
1827                 if (!xas_error(&xas))
1828                         break;
1829                 xas_unlock_irq(&xas);
1830                 if (!xas_nomem(&xas, GFP_KERNEL)) {
1831                         result = SCAN_FAIL;
1832                         goto rollback;
1833                 }
1834         } while (1);
1835
1836         for (index = start; index < end; index++) {
1837                 xas_set(&xas, index);
1838                 page = xas_load(&xas);
1839
1840                 VM_BUG_ON(index != xas.xa_index);
1841                 if (is_shmem) {
1842                         if (!page) {
1843                                 /*
1844                                  * Stop if extent has been truncated or
1845                                  * hole-punched, and is now completely
1846                                  * empty.
1847                                  */
1848                                 if (index == start) {
1849                                         if (!xas_next_entry(&xas, end - 1)) {
1850                                                 result = SCAN_TRUNCATED;
1851                                                 goto xa_locked;
1852                                         }
1853                                 }
1854                                 nr_none++;
1855                                 continue;
1856                         }
1857
1858                         if (xa_is_value(page) || !PageUptodate(page)) {
1859                                 xas_unlock_irq(&xas);
1860                                 /* swap in or instantiate fallocated page */
1861                                 if (shmem_get_folio(mapping->host, index,
1862                                                 &folio, SGP_NOALLOC)) {
1863                                         result = SCAN_FAIL;
1864                                         goto xa_unlocked;
1865                                 }
1866                                 /* drain lru cache to help isolate_lru_page() */
1867                                 lru_add_drain();
1868                                 page = folio_file_page(folio, index);
1869                         } else if (trylock_page(page)) {
1870                                 get_page(page);
1871                                 xas_unlock_irq(&xas);
1872                         } else {
1873                                 result = SCAN_PAGE_LOCK;
1874                                 goto xa_locked;
1875                         }
1876                 } else {        /* !is_shmem */
1877                         if (!page || xa_is_value(page)) {
1878                                 xas_unlock_irq(&xas);
1879                                 page_cache_sync_readahead(mapping, &file->f_ra,
1880                                                           file, index,
1881                                                           end - index);
1882                                 /* drain lru cache to help isolate_lru_page() */
1883                                 lru_add_drain();
1884                                 page = find_lock_page(mapping, index);
1885                                 if (unlikely(page == NULL)) {
1886                                         result = SCAN_FAIL;
1887                                         goto xa_unlocked;
1888                                 }
1889                         } else if (PageDirty(page)) {
1890                                 /*
1891                                  * khugepaged only works on read-only fd,
1892                                  * so this page is dirty because it hasn't
1893                                  * been flushed since first write. There
1894                                  * won't be new dirty pages.
1895                                  *
1896                                  * Trigger async flush here and hope the
1897                                  * writeback is done when khugepaged
1898                                  * revisits this page.
1899                                  *
1900                                  * This is a one-off situation. We are not
1901                                  * forcing writeback in loop.
1902                                  */
1903                                 xas_unlock_irq(&xas);
1904                                 filemap_flush(mapping);
1905                                 result = SCAN_FAIL;
1906                                 goto xa_unlocked;
1907                         } else if (PageWriteback(page)) {
1908                                 xas_unlock_irq(&xas);
1909                                 result = SCAN_FAIL;
1910                                 goto xa_unlocked;
1911                         } else if (trylock_page(page)) {
1912                                 get_page(page);
1913                                 xas_unlock_irq(&xas);
1914                         } else {
1915                                 result = SCAN_PAGE_LOCK;
1916                                 goto xa_locked;
1917                         }
1918                 }
1919
1920                 /*
1921                  * The page must be locked, so we can drop the i_pages lock
1922                  * without racing with truncate.
1923                  */
1924                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1925
1926                 /* make sure the page is up to date */
1927                 if (unlikely(!PageUptodate(page))) {
1928                         result = SCAN_FAIL;
1929                         goto out_unlock;
1930                 }
1931
1932                 /*
1933                  * If file was truncated then extended, or hole-punched, before
1934                  * we locked the first page, then a THP might be there already.
1935                  * This will be discovered on the first iteration.
1936                  */
1937                 if (PageTransCompound(page)) {
1938                         struct page *head = compound_head(page);
1939
1940                         result = compound_order(head) == HPAGE_PMD_ORDER &&
1941                                         head->index == start
1942                                         /* Maybe PMD-mapped */
1943                                         ? SCAN_PTE_MAPPED_HUGEPAGE
1944                                         : SCAN_PAGE_COMPOUND;
1945                         goto out_unlock;
1946                 }
1947
1948                 folio = page_folio(page);
1949
1950                 if (folio_mapping(folio) != mapping) {
1951                         result = SCAN_TRUNCATED;
1952                         goto out_unlock;
1953                 }
1954
1955                 if (!is_shmem && (folio_test_dirty(folio) ||
1956                                   folio_test_writeback(folio))) {
1957                         /*
1958                          * khugepaged only works on read-only fd, so this
1959                          * page is dirty because it hasn't been flushed
1960                          * since first write.
1961                          */
1962                         result = SCAN_FAIL;
1963                         goto out_unlock;
1964                 }
1965
1966                 if (!folio_isolate_lru(folio)) {
1967                         result = SCAN_DEL_PAGE_LRU;
1968                         goto out_unlock;
1969                 }
1970
1971                 if (!filemap_release_folio(folio, GFP_KERNEL)) {
1972                         result = SCAN_PAGE_HAS_PRIVATE;
1973                         folio_putback_lru(folio);
1974                         goto out_unlock;
1975                 }
1976
1977                 if (folio_mapped(folio))
1978                         try_to_unmap(folio,
1979                                         TTU_IGNORE_MLOCK | TTU_BATCH_FLUSH);
1980
1981                 xas_lock_irq(&xas);
1982
1983                 VM_BUG_ON_PAGE(page != xa_load(xas.xa, index), page);
1984
1985                 /*
1986                  * We control three references to the page:
1987                  *  - we hold a pin on it;
1988                  *  - one reference from page cache;
1989                  *  - one from isolate_lru_page;
1990                  * If those are the only references, then any new usage of the
1991                  * page will have to fetch it from the page cache. That requires
1992                  * locking the page to handle truncate, so any new usage will be
1993                  * blocked until we unlock page after collapse/during rollback.
1994                  */
1995                 if (page_count(page) != 3) {
1996                         result = SCAN_PAGE_COUNT;
1997                         xas_unlock_irq(&xas);
1998                         putback_lru_page(page);
1999                         goto out_unlock;
2000                 }
2001
2002                 /*
2003                  * Accumulate the pages that are being collapsed.
2004                  */
2005                 list_add_tail(&page->lru, &pagelist);
2006                 continue;
2007 out_unlock:
2008                 unlock_page(page);
2009                 put_page(page);
2010                 goto xa_unlocked;
2011         }
2012
2013         if (!is_shmem) {
2014                 filemap_nr_thps_inc(mapping);
2015                 /*
2016                  * Paired with smp_mb() in do_dentry_open() to ensure
2017                  * i_writecount is up to date and the update to nr_thps is
2018                  * visible. Ensures the page cache will be truncated if the
2019                  * file is opened writable.
2020                  */
2021                 smp_mb();
2022                 if (inode_is_open_for_write(mapping->host)) {
2023                         result = SCAN_FAIL;
2024                         filemap_nr_thps_dec(mapping);
2025                 }
2026         }
2027
2028 xa_locked:
2029         xas_unlock_irq(&xas);
2030 xa_unlocked:
2031
2032         /*
2033          * If collapse is successful, flush must be done now before copying.
2034          * If collapse is unsuccessful, does flush actually need to be done?
2035          * Do it anyway, to clear the state.
2036          */
2037         try_to_unmap_flush();
2038
2039         if (result == SCAN_SUCCEED && nr_none &&
2040             !shmem_charge(mapping->host, nr_none))
2041                 result = SCAN_FAIL;
2042         if (result != SCAN_SUCCEED) {
2043                 nr_none = 0;
2044                 goto rollback;
2045         }
2046
2047         /*
2048          * The old pages are locked, so they won't change anymore.
2049          */
2050         index = start;
2051         list_for_each_entry(page, &pagelist, lru) {
2052                 while (index < page->index) {
2053                         clear_highpage(hpage + (index % HPAGE_PMD_NR));
2054                         index++;
2055                 }
2056                 if (copy_mc_highpage(hpage + (page->index % HPAGE_PMD_NR), page) > 0) {
2057                         result = SCAN_COPY_MC;
2058                         goto rollback;
2059                 }
2060                 index++;
2061         }
2062         while (index < end) {
2063                 clear_highpage(hpage + (index % HPAGE_PMD_NR));
2064                 index++;
2065         }
2066
2067         if (nr_none) {
2068                 struct vm_area_struct *vma;
2069                 int nr_none_check = 0;
2070
2071                 i_mmap_lock_read(mapping);
2072                 xas_lock_irq(&xas);
2073
2074                 xas_set(&xas, start);
2075                 for (index = start; index < end; index++) {
2076                         if (!xas_next(&xas)) {
2077                                 xas_store(&xas, XA_RETRY_ENTRY);
2078                                 if (xas_error(&xas)) {
2079                                         result = SCAN_STORE_FAILED;
2080                                         goto immap_locked;
2081                                 }
2082                                 nr_none_check++;
2083                         }
2084                 }
2085
2086                 if (nr_none != nr_none_check) {
2087                         result = SCAN_PAGE_FILLED;
2088                         goto immap_locked;
2089                 }
2090
2091                 /*
2092                  * If userspace observed a missing page in a VMA with a MODE_MISSING
2093                  * userfaultfd, then it might expect a UFFD_EVENT_PAGEFAULT for that
2094                  * page. If so, we need to roll back to avoid suppressing such an
2095                  * event. Since wp/minor userfaultfds don't give userspace any
2096                  * guarantees that the kernel doesn't fill a missing page with a zero
2097                  * page, so they don't matter here.
2098                  *
2099                  * Any userfaultfds registered after this point will not be able to
2100                  * observe any missing pages due to the previously inserted retry
2101                  * entries.
2102                  */
2103                 vma_interval_tree_foreach(vma, &mapping->i_mmap, start, end) {
2104                         if (userfaultfd_missing(vma)) {
2105                                 result = SCAN_EXCEED_NONE_PTE;
2106                                 goto immap_locked;
2107                         }
2108                 }
2109
2110 immap_locked:
2111                 i_mmap_unlock_read(mapping);
2112                 if (result != SCAN_SUCCEED) {
2113                         xas_set(&xas, start);
2114                         for (index = start; index < end; index++) {
2115                                 if (xas_next(&xas) == XA_RETRY_ENTRY)
2116                                         xas_store(&xas, NULL);
2117                         }
2118
2119                         xas_unlock_irq(&xas);
2120                         goto rollback;
2121                 }
2122         } else {
2123                 xas_lock_irq(&xas);
2124         }
2125
2126         folio = page_folio(hpage);
2127         nr = folio_nr_pages(folio);
2128         if (is_shmem)
2129                 __lruvec_stat_mod_folio(folio, NR_SHMEM_THPS, nr);
2130         else
2131                 __lruvec_stat_mod_folio(folio, NR_FILE_THPS, nr);
2132
2133         if (nr_none) {
2134                 __lruvec_stat_mod_folio(folio, NR_FILE_PAGES, nr_none);
2135                 /* nr_none is always 0 for non-shmem. */
2136                 __lruvec_stat_mod_folio(folio, NR_SHMEM, nr_none);
2137         }
2138
2139         /*
2140          * Mark hpage as uptodate before inserting it into the page cache so
2141          * that it isn't mistaken for an fallocated but unwritten page.
2142          */
2143         folio_mark_uptodate(folio);
2144         folio_ref_add(folio, HPAGE_PMD_NR - 1);
2145
2146         if (is_shmem)
2147                 folio_mark_dirty(folio);
2148         folio_add_lru(folio);
2149
2150         /* Join all the small entries into a single multi-index entry. */
2151         xas_set_order(&xas, start, HPAGE_PMD_ORDER);
2152         xas_store(&xas, folio);
2153         WARN_ON_ONCE(xas_error(&xas));
2154         xas_unlock_irq(&xas);
2155
2156         /*
2157          * Remove pte page tables, so we can re-fault the page as huge.
2158          * If MADV_COLLAPSE, adjust result to call collapse_pte_mapped_thp().
2159          */
2160         retract_page_tables(mapping, start);
2161         if (cc && !cc->is_khugepaged)
2162                 result = SCAN_PTE_MAPPED_HUGEPAGE;
2163         folio_unlock(folio);
2164
2165         /*
2166          * The collapse has succeeded, so free the old pages.
2167          */
2168         list_for_each_entry_safe(page, tmp, &pagelist, lru) {
2169                 list_del(&page->lru);
2170                 page->mapping = NULL;
2171                 ClearPageActive(page);
2172                 ClearPageUnevictable(page);
2173                 unlock_page(page);
2174                 folio_put_refs(page_folio(page), 3);
2175         }
2176
2177         goto out;
2178
2179 rollback:
2180         /* Something went wrong: roll back page cache changes */
2181         if (nr_none) {
2182                 xas_lock_irq(&xas);
2183                 mapping->nrpages -= nr_none;
2184                 xas_unlock_irq(&xas);
2185                 shmem_uncharge(mapping->host, nr_none);
2186         }
2187
2188         list_for_each_entry_safe(page, tmp, &pagelist, lru) {
2189                 list_del(&page->lru);
2190                 unlock_page(page);
2191                 putback_lru_page(page);
2192                 put_page(page);
2193         }
2194         /*
2195          * Undo the updates of filemap_nr_thps_inc for non-SHMEM
2196          * file only. This undo is not needed unless failure is
2197          * due to SCAN_COPY_MC.
2198          */
2199         if (!is_shmem && result == SCAN_COPY_MC) {
2200                 filemap_nr_thps_dec(mapping);
2201                 /*
2202                  * Paired with smp_mb() in do_dentry_open() to
2203                  * ensure the update to nr_thps is visible.
2204                  */
2205                 smp_mb();
2206         }
2207
2208         hpage->mapping = NULL;
2209
2210         unlock_page(hpage);
2211         put_page(hpage);
2212 out:
2213         VM_BUG_ON(!list_empty(&pagelist));
2214         trace_mm_khugepaged_collapse_file(mm, hpage, index, is_shmem, addr, file, nr, result);
2215         return result;
2216 }
2217
2218 static int hpage_collapse_scan_file(struct mm_struct *mm, unsigned long addr,
2219                                     struct file *file, pgoff_t start,
2220                                     struct collapse_control *cc)
2221 {
2222         struct page *page = NULL;
2223         struct address_space *mapping = file->f_mapping;
2224         XA_STATE(xas, &mapping->i_pages, start);
2225         int present, swap;
2226         int node = NUMA_NO_NODE;
2227         int result = SCAN_SUCCEED;
2228
2229         present = 0;
2230         swap = 0;
2231         memset(cc->node_load, 0, sizeof(cc->node_load));
2232         nodes_clear(cc->alloc_nmask);
2233         rcu_read_lock();
2234         xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
2235                 if (xas_retry(&xas, page))
2236                         continue;
2237
2238                 if (xa_is_value(page)) {
2239                         ++swap;
2240                         if (cc->is_khugepaged &&
2241                             swap > khugepaged_max_ptes_swap) {
2242                                 result = SCAN_EXCEED_SWAP_PTE;
2243                                 count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
2244                                 break;
2245                         }
2246                         continue;
2247                 }
2248
2249                 /*
2250                  * TODO: khugepaged should compact smaller compound pages
2251                  * into a PMD sized page
2252                  */
2253                 if (PageTransCompound(page)) {
2254                         struct page *head = compound_head(page);
2255
2256                         result = compound_order(head) == HPAGE_PMD_ORDER &&
2257                                         head->index == start
2258                                         /* Maybe PMD-mapped */
2259                                         ? SCAN_PTE_MAPPED_HUGEPAGE
2260                                         : SCAN_PAGE_COMPOUND;
2261                         /*
2262                          * For SCAN_PTE_MAPPED_HUGEPAGE, further processing
2263                          * by the caller won't touch the page cache, and so
2264                          * it's safe to skip LRU and refcount checks before
2265                          * returning.
2266                          */
2267                         break;
2268                 }
2269
2270                 node = page_to_nid(page);
2271                 if (hpage_collapse_scan_abort(node, cc)) {
2272                         result = SCAN_SCAN_ABORT;
2273                         break;
2274                 }
2275                 cc->node_load[node]++;
2276
2277                 if (!PageLRU(page)) {
2278                         result = SCAN_PAGE_LRU;
2279                         break;
2280                 }
2281
2282                 if (page_count(page) !=
2283                     1 + page_mapcount(page) + page_has_private(page)) {
2284                         result = SCAN_PAGE_COUNT;
2285                         break;
2286                 }
2287
2288                 /*
2289                  * We probably should check if the page is referenced here, but
2290                  * nobody would transfer pte_young() to PageReferenced() for us.
2291                  * And rmap walk here is just too costly...
2292                  */
2293
2294                 present++;
2295
2296                 if (need_resched()) {
2297                         xas_pause(&xas);
2298                         cond_resched_rcu();
2299                 }
2300         }
2301         rcu_read_unlock();
2302
2303         if (result == SCAN_SUCCEED) {
2304                 if (cc->is_khugepaged &&
2305                     present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2306                         result = SCAN_EXCEED_NONE_PTE;
2307                         count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
2308                 } else {
2309                         result = collapse_file(mm, addr, file, start, cc);
2310                 }
2311         }
2312
2313         trace_mm_khugepaged_scan_file(mm, page, file, present, swap, result);
2314         return result;
2315 }
2316 #else
2317 static int hpage_collapse_scan_file(struct mm_struct *mm, unsigned long addr,
2318                                     struct file *file, pgoff_t start,
2319                                     struct collapse_control *cc)
2320 {
2321         BUILD_BUG();
2322 }
2323 #endif
2324
2325 static unsigned int khugepaged_scan_mm_slot(unsigned int pages, int *result,
2326                                             struct collapse_control *cc)
2327         __releases(&khugepaged_mm_lock)
2328         __acquires(&khugepaged_mm_lock)
2329 {
2330         struct vma_iterator vmi;
2331         struct khugepaged_mm_slot *mm_slot;
2332         struct mm_slot *slot;
2333         struct mm_struct *mm;
2334         struct vm_area_struct *vma;
2335         int progress = 0;
2336
2337         VM_BUG_ON(!pages);
2338         lockdep_assert_held(&khugepaged_mm_lock);
2339         *result = SCAN_FAIL;
2340
2341         if (khugepaged_scan.mm_slot) {
2342                 mm_slot = khugepaged_scan.mm_slot;
2343                 slot = &mm_slot->slot;
2344         } else {
2345                 slot = list_entry(khugepaged_scan.mm_head.next,
2346                                      struct mm_slot, mm_node);
2347                 mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot);
2348                 khugepaged_scan.address = 0;
2349                 khugepaged_scan.mm_slot = mm_slot;
2350         }
2351         spin_unlock(&khugepaged_mm_lock);
2352
2353         mm = slot->mm;
2354         /*
2355          * Don't wait for semaphore (to avoid long wait times).  Just move to
2356          * the next mm on the list.
2357          */
2358         vma = NULL;
2359         if (unlikely(!mmap_read_trylock(mm)))
2360                 goto breakouterloop_mmap_lock;
2361
2362         progress++;
2363         if (unlikely(hpage_collapse_test_exit(mm)))
2364                 goto breakouterloop;
2365
2366         vma_iter_init(&vmi, mm, khugepaged_scan.address);
2367         for_each_vma(vmi, vma) {
2368                 unsigned long hstart, hend;
2369
2370                 cond_resched();
2371                 if (unlikely(hpage_collapse_test_exit(mm))) {
2372                         progress++;
2373                         break;
2374                 }
2375                 if (!thp_vma_allowable_order(vma, vma->vm_flags, false, false,
2376                                              true, PMD_ORDER)) {
2377 skip:
2378                         progress++;
2379                         continue;
2380                 }
2381                 hstart = round_up(vma->vm_start, HPAGE_PMD_SIZE);
2382                 hend = round_down(vma->vm_end, HPAGE_PMD_SIZE);
2383                 if (khugepaged_scan.address > hend)
2384                         goto skip;
2385                 if (khugepaged_scan.address < hstart)
2386                         khugepaged_scan.address = hstart;
2387                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2388
2389                 while (khugepaged_scan.address < hend) {
2390                         bool mmap_locked = true;
2391
2392                         cond_resched();
2393                         if (unlikely(hpage_collapse_test_exit(mm)))
2394                                 goto breakouterloop;
2395
2396                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2397                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2398                                   hend);
2399                         if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2400                                 struct file *file = get_file(vma->vm_file);
2401                                 pgoff_t pgoff = linear_page_index(vma,
2402                                                 khugepaged_scan.address);
2403
2404                                 mmap_read_unlock(mm);
2405                                 mmap_locked = false;
2406                                 *result = hpage_collapse_scan_file(mm,
2407                                         khugepaged_scan.address, file, pgoff, cc);
2408                                 fput(file);
2409                                 if (*result == SCAN_PTE_MAPPED_HUGEPAGE) {
2410                                         mmap_read_lock(mm);
2411                                         if (hpage_collapse_test_exit(mm))
2412                                                 goto breakouterloop;
2413                                         *result = collapse_pte_mapped_thp(mm,
2414                                                 khugepaged_scan.address, false);
2415                                         if (*result == SCAN_PMD_MAPPED)
2416                                                 *result = SCAN_SUCCEED;
2417                                         mmap_read_unlock(mm);
2418                                 }
2419                         } else {
2420                                 *result = hpage_collapse_scan_pmd(mm, vma,
2421                                         khugepaged_scan.address, &mmap_locked, cc);
2422                         }
2423
2424                         if (*result == SCAN_SUCCEED)
2425                                 ++khugepaged_pages_collapsed;
2426
2427                         /* move to next address */
2428                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2429                         progress += HPAGE_PMD_NR;
2430                         if (!mmap_locked)
2431                                 /*
2432                                  * We released mmap_lock so break loop.  Note
2433                                  * that we drop mmap_lock before all hugepage
2434                                  * allocations, so if allocation fails, we are
2435                                  * guaranteed to break here and report the
2436                                  * correct result back to caller.
2437                                  */
2438                                 goto breakouterloop_mmap_lock;
2439                         if (progress >= pages)
2440                                 goto breakouterloop;
2441                 }
2442         }
2443 breakouterloop:
2444         mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2445 breakouterloop_mmap_lock:
2446
2447         spin_lock(&khugepaged_mm_lock);
2448         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2449         /*
2450          * Release the current mm_slot if this mm is about to die, or
2451          * if we scanned all vmas of this mm.
2452          */
2453         if (hpage_collapse_test_exit(mm) || !vma) {
2454                 /*
2455                  * Make sure that if mm_users is reaching zero while
2456                  * khugepaged runs here, khugepaged_exit will find
2457                  * mm_slot not pointing to the exiting mm.
2458                  */
2459                 if (slot->mm_node.next != &khugepaged_scan.mm_head) {
2460                         slot = list_entry(slot->mm_node.next,
2461                                           struct mm_slot, mm_node);
2462                         khugepaged_scan.mm_slot =
2463                                 mm_slot_entry(slot, struct khugepaged_mm_slot, slot);
2464                         khugepaged_scan.address = 0;
2465                 } else {
2466                         khugepaged_scan.mm_slot = NULL;
2467                         khugepaged_full_scans++;
2468                 }
2469
2470                 collect_mm_slot(mm_slot);
2471         }
2472
2473         return progress;
2474 }
2475
2476 static int khugepaged_has_work(void)
2477 {
2478         return !list_empty(&khugepaged_scan.mm_head) &&
2479                 hugepage_flags_enabled();
2480 }
2481
2482 static int khugepaged_wait_event(void)
2483 {
2484         return !list_empty(&khugepaged_scan.mm_head) ||
2485                 kthread_should_stop();
2486 }
2487
2488 static void khugepaged_do_scan(struct collapse_control *cc)
2489 {
2490         unsigned int progress = 0, pass_through_head = 0;
2491         unsigned int pages = READ_ONCE(khugepaged_pages_to_scan);
2492         bool wait = true;
2493         int result = SCAN_SUCCEED;
2494
2495         lru_add_drain_all();
2496
2497         while (true) {
2498                 cond_resched();
2499
2500                 if (unlikely(kthread_should_stop()))
2501                         break;
2502
2503                 spin_lock(&khugepaged_mm_lock);
2504                 if (!khugepaged_scan.mm_slot)
2505                         pass_through_head++;
2506                 if (khugepaged_has_work() &&
2507                     pass_through_head < 2)
2508                         progress += khugepaged_scan_mm_slot(pages - progress,
2509                                                             &result, cc);
2510                 else
2511                         progress = pages;
2512                 spin_unlock(&khugepaged_mm_lock);
2513
2514                 if (progress >= pages)
2515                         break;
2516
2517                 if (result == SCAN_ALLOC_HUGE_PAGE_FAIL) {
2518                         /*
2519                          * If fail to allocate the first time, try to sleep for
2520                          * a while.  When hit again, cancel the scan.
2521                          */
2522                         if (!wait)
2523                                 break;
2524                         wait = false;
2525                         khugepaged_alloc_sleep();
2526                 }
2527         }
2528 }
2529
2530 static bool khugepaged_should_wakeup(void)
2531 {
2532         return kthread_should_stop() ||
2533                time_after_eq(jiffies, khugepaged_sleep_expire);
2534 }
2535
2536 static void khugepaged_wait_work(void)
2537 {
2538         if (khugepaged_has_work()) {
2539                 const unsigned long scan_sleep_jiffies =
2540                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2541
2542                 if (!scan_sleep_jiffies)
2543                         return;
2544
2545                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2546                 wait_event_freezable_timeout(khugepaged_wait,
2547                                              khugepaged_should_wakeup(),
2548                                              scan_sleep_jiffies);
2549                 return;
2550         }
2551
2552         if (hugepage_flags_enabled())
2553                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2554 }
2555
2556 static int khugepaged(void *none)
2557 {
2558         struct khugepaged_mm_slot *mm_slot;
2559
2560         set_freezable();
2561         set_user_nice(current, MAX_NICE);
2562
2563         while (!kthread_should_stop()) {
2564                 khugepaged_do_scan(&khugepaged_collapse_control);
2565                 khugepaged_wait_work();
2566         }
2567
2568         spin_lock(&khugepaged_mm_lock);
2569         mm_slot = khugepaged_scan.mm_slot;
2570         khugepaged_scan.mm_slot = NULL;
2571         if (mm_slot)
2572                 collect_mm_slot(mm_slot);
2573         spin_unlock(&khugepaged_mm_lock);
2574         return 0;
2575 }
2576
2577 static void set_recommended_min_free_kbytes(void)
2578 {
2579         struct zone *zone;
2580         int nr_zones = 0;
2581         unsigned long recommended_min;
2582
2583         if (!hugepage_flags_enabled()) {
2584                 calculate_min_free_kbytes();
2585                 goto update_wmarks;
2586         }
2587
2588         for_each_populated_zone(zone) {
2589                 /*
2590                  * We don't need to worry about fragmentation of
2591                  * ZONE_MOVABLE since it only has movable pages.
2592                  */
2593                 if (zone_idx(zone) > gfp_zone(GFP_USER))
2594                         continue;
2595
2596                 nr_zones++;
2597         }
2598
2599         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2600         recommended_min = pageblock_nr_pages * nr_zones * 2;
2601
2602         /*
2603          * Make sure that on average at least two pageblocks are almost free
2604          * of another type, one for a migratetype to fall back to and a
2605          * second to avoid subsequent fallbacks of other types There are 3
2606          * MIGRATE_TYPES we care about.
2607          */
2608         recommended_min += pageblock_nr_pages * nr_zones *
2609                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2610
2611         /* don't ever allow to reserve more than 5% of the lowmem */
2612         recommended_min = min(recommended_min,
2613                               (unsigned long) nr_free_buffer_pages() / 20);
2614         recommended_min <<= (PAGE_SHIFT-10);
2615
2616         if (recommended_min > min_free_kbytes) {
2617                 if (user_min_free_kbytes >= 0)
2618                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2619                                 min_free_kbytes, recommended_min);
2620
2621                 min_free_kbytes = recommended_min;
2622         }
2623
2624 update_wmarks:
2625         setup_per_zone_wmarks();
2626 }
2627
2628 int start_stop_khugepaged(void)
2629 {
2630         int err = 0;
2631
2632         mutex_lock(&khugepaged_mutex);
2633         if (hugepage_flags_enabled()) {
2634                 if (!khugepaged_thread)
2635                         khugepaged_thread = kthread_run(khugepaged, NULL,
2636                                                         "khugepaged");
2637                 if (IS_ERR(khugepaged_thread)) {
2638                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2639                         err = PTR_ERR(khugepaged_thread);
2640                         khugepaged_thread = NULL;
2641                         goto fail;
2642                 }
2643
2644                 if (!list_empty(&khugepaged_scan.mm_head))
2645                         wake_up_interruptible(&khugepaged_wait);
2646         } else if (khugepaged_thread) {
2647                 kthread_stop(khugepaged_thread);
2648                 khugepaged_thread = NULL;
2649         }
2650         set_recommended_min_free_kbytes();
2651 fail:
2652         mutex_unlock(&khugepaged_mutex);
2653         return err;
2654 }
2655
2656 void khugepaged_min_free_kbytes_update(void)
2657 {
2658         mutex_lock(&khugepaged_mutex);
2659         if (hugepage_flags_enabled() && khugepaged_thread)
2660                 set_recommended_min_free_kbytes();
2661         mutex_unlock(&khugepaged_mutex);
2662 }
2663
2664 bool current_is_khugepaged(void)
2665 {
2666         return kthread_func(current) == khugepaged;
2667 }
2668
2669 static int madvise_collapse_errno(enum scan_result r)
2670 {
2671         /*
2672          * MADV_COLLAPSE breaks from existing madvise(2) conventions to provide
2673          * actionable feedback to caller, so they may take an appropriate
2674          * fallback measure depending on the nature of the failure.
2675          */
2676         switch (r) {
2677         case SCAN_ALLOC_HUGE_PAGE_FAIL:
2678                 return -ENOMEM;
2679         case SCAN_CGROUP_CHARGE_FAIL:
2680         case SCAN_EXCEED_NONE_PTE:
2681                 return -EBUSY;
2682         /* Resource temporary unavailable - trying again might succeed */
2683         case SCAN_PAGE_COUNT:
2684         case SCAN_PAGE_LOCK:
2685         case SCAN_PAGE_LRU:
2686         case SCAN_DEL_PAGE_LRU:
2687         case SCAN_PAGE_FILLED:
2688                 return -EAGAIN;
2689         /*
2690          * Other: Trying again likely not to succeed / error intrinsic to
2691          * specified memory range. khugepaged likely won't be able to collapse
2692          * either.
2693          */
2694         default:
2695                 return -EINVAL;
2696         }
2697 }
2698
2699 int madvise_collapse(struct vm_area_struct *vma, struct vm_area_struct **prev,
2700                      unsigned long start, unsigned long end)
2701 {
2702         struct collapse_control *cc;
2703         struct mm_struct *mm = vma->vm_mm;
2704         unsigned long hstart, hend, addr;
2705         int thps = 0, last_fail = SCAN_FAIL;
2706         bool mmap_locked = true;
2707
2708         BUG_ON(vma->vm_start > start);
2709         BUG_ON(vma->vm_end < end);
2710
2711         *prev = vma;
2712
2713         if (!thp_vma_allowable_order(vma, vma->vm_flags, false, false, false,
2714                                      PMD_ORDER))
2715                 return -EINVAL;
2716
2717         cc = kmalloc(sizeof(*cc), GFP_KERNEL);
2718         if (!cc)
2719                 return -ENOMEM;
2720         cc->is_khugepaged = false;
2721
2722         mmgrab(mm);
2723         lru_add_drain_all();
2724
2725         hstart = (start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2726         hend = end & HPAGE_PMD_MASK;
2727
2728         for (addr = hstart; addr < hend; addr += HPAGE_PMD_SIZE) {
2729                 int result = SCAN_FAIL;
2730
2731                 if (!mmap_locked) {
2732                         cond_resched();
2733                         mmap_read_lock(mm);
2734                         mmap_locked = true;
2735                         result = hugepage_vma_revalidate(mm, addr, false, &vma,
2736                                                          cc);
2737                         if (result  != SCAN_SUCCEED) {
2738                                 last_fail = result;
2739                                 goto out_nolock;
2740                         }
2741
2742                         hend = min(hend, vma->vm_end & HPAGE_PMD_MASK);
2743                 }
2744                 mmap_assert_locked(mm);
2745                 memset(cc->node_load, 0, sizeof(cc->node_load));
2746                 nodes_clear(cc->alloc_nmask);
2747                 if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2748                         struct file *file = get_file(vma->vm_file);
2749                         pgoff_t pgoff = linear_page_index(vma, addr);
2750
2751                         mmap_read_unlock(mm);
2752                         mmap_locked = false;
2753                         result = hpage_collapse_scan_file(mm, addr, file, pgoff,
2754                                                           cc);
2755                         fput(file);
2756                 } else {
2757                         result = hpage_collapse_scan_pmd(mm, vma, addr,
2758                                                          &mmap_locked, cc);
2759                 }
2760                 if (!mmap_locked)
2761                         *prev = NULL;  /* Tell caller we dropped mmap_lock */
2762
2763 handle_result:
2764                 switch (result) {
2765                 case SCAN_SUCCEED:
2766                 case SCAN_PMD_MAPPED:
2767                         ++thps;
2768                         break;
2769                 case SCAN_PTE_MAPPED_HUGEPAGE:
2770                         BUG_ON(mmap_locked);
2771                         BUG_ON(*prev);
2772                         mmap_read_lock(mm);
2773                         result = collapse_pte_mapped_thp(mm, addr, true);
2774                         mmap_read_unlock(mm);
2775                         goto handle_result;
2776                 /* Whitelisted set of results where continuing OK */
2777                 case SCAN_PMD_NULL:
2778                 case SCAN_PTE_NON_PRESENT:
2779                 case SCAN_PTE_UFFD_WP:
2780                 case SCAN_PAGE_RO:
2781                 case SCAN_LACK_REFERENCED_PAGE:
2782                 case SCAN_PAGE_NULL:
2783                 case SCAN_PAGE_COUNT:
2784                 case SCAN_PAGE_LOCK:
2785                 case SCAN_PAGE_COMPOUND:
2786                 case SCAN_PAGE_LRU:
2787                 case SCAN_DEL_PAGE_LRU:
2788                         last_fail = result;
2789                         break;
2790                 default:
2791                         last_fail = result;
2792                         /* Other error, exit */
2793                         goto out_maybelock;
2794                 }
2795         }
2796
2797 out_maybelock:
2798         /* Caller expects us to hold mmap_lock on return */
2799         if (!mmap_locked)
2800                 mmap_read_lock(mm);
2801 out_nolock:
2802         mmap_assert_locked(mm);
2803         mmdrop(mm);
2804         kfree(cc);
2805
2806         return thps == ((hend - hstart) >> HPAGE_PMD_SHIFT) ? 0
2807                         : madvise_collapse_errno(last_fail);
2808 }