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(khugepaged_mm_slot, 0);
 389         if (!mm_slot_cache)
 390                 return -ENOMEM;
 391
 392         khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
 393         khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
 394         khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
 395         khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2;
 396
 397         return 0;
 398 }
 399
 400 void __init khugepaged_destroy(void)
 401 {
 402         kmem_cache_destroy(mm_slot_cache);
 403 }
 404
 405 static inline int hpage_collapse_test_exit(struct mm_struct *mm)
 406 {
 407         return atomic_read(&mm->mm_users) == 0;
 408 }
 409
 410 static inline int hpage_collapse_test_exit_or_disable(struct mm_struct *mm)
 411 {
 412         return hpage_collapse_test_exit(mm) ||
 413                test_bit(MMF_DISABLE_THP, &mm->flags);
 414 }
 415
 416 static bool hugepage_pmd_enabled(void)
 417 {
 418         /*
 419          * We cover both the anon and the file-backed case here; file-backed
 420          * hugepages, when configured in, are determined by the global control.
 421          * Anon pmd-sized hugepages are determined by the pmd-size control.
 422          */
 423         if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) &&
 424             hugepage_global_enabled())
 425                 return true;
 426         if (test_bit(PMD_ORDER, &huge_anon_orders_always))
 427                 return true;
 428         if (test_bit(PMD_ORDER, &huge_anon_orders_madvise))
 429                 return true;
 430         if (test_bit(PMD_ORDER, &huge_anon_orders_inherit) &&
 431             hugepage_global_enabled())
 432                 return true;
 433         return false;
 434 }
 435
 436 void __khugepaged_enter(struct mm_struct *mm)
 437 {
 438         struct khugepaged_mm_slot *mm_slot;
 439         struct mm_slot *slot;
 440         int wakeup;
 441
 442         /* __khugepaged_exit() must not run from under us */
 443         VM_BUG_ON_MM(hpage_collapse_test_exit(mm), mm);
 444         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags)))
 445                 return;
 446
 447         mm_slot = mm_slot_alloc(mm_slot_cache);
 448         if (!mm_slot)
 449                 return;
 450
 451         slot = &mm_slot->slot;
 452
 453         spin_lock(&khugepaged_mm_lock);
 454         mm_slot_insert(mm_slots_hash, mm, slot);
 455         /*
 456          * Insert just behind the scanning cursor, to let the area settle
 457          * down a little.
 458          */
 459         wakeup = list_empty(&khugepaged_scan.mm_head);
 460         list_add_tail(&slot->mm_node, &khugepaged_scan.mm_head);
 461         spin_unlock(&khugepaged_mm_lock);
 462
 463         mmgrab(mm);
 464         if (wakeup)
 465                 wake_up_interruptible(&khugepaged_wait);
 466 }
 467
 468 void khugepaged_enter_vma(struct vm_area_struct *vma,
 469                           unsigned long vm_flags)
 470 {
 471         if (!test_bit(MMF_VM_HUGEPAGE, &vma->vm_mm->flags) &&
 472             hugepage_pmd_enabled()) {
 473                 if (thp_vma_allowable_order(vma, vm_flags, TVA_ENFORCE_SYSFS,
 474                                             PMD_ORDER))
 475                         __khugepaged_enter(vma->vm_mm);
 476         }
 477 }
 478
 479 void __khugepaged_exit(struct mm_struct *mm)
 480 {
 481         struct khugepaged_mm_slot *mm_slot;
 482         struct mm_slot *slot;
 483         int free = 0;
 484
 485         spin_lock(&khugepaged_mm_lock);
 486         slot = mm_slot_lookup(mm_slots_hash, mm);
 487         mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot);
 488         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
 489                 hash_del(&slot->hash);
 490                 list_del(&slot->mm_node);
 491                 free = 1;
 492         }
 493         spin_unlock(&khugepaged_mm_lock);
 494
 495         if (free) {
 496                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
 497                 mm_slot_free(mm_slot_cache, mm_slot);
 498                 mmdrop(mm);
 499         } else if (mm_slot) {
 500                 /*
 501                  * This is required to serialize against
 502                  * hpage_collapse_test_exit() (which is guaranteed to run
 503                  * under mmap sem read mode). Stop here (after we return all
 504                  * pagetables will be destroyed) until khugepaged has finished
 505                  * working on the pagetables under the mmap_lock.
 506                  */
 507                 mmap_write_lock(mm);
 508                 mmap_write_unlock(mm);
 509         }
 510 }
 511
 512 static void release_pte_folio(struct folio *folio)
 513 {
 514         node_stat_mod_folio(folio,
 515                         NR_ISOLATED_ANON + folio_is_file_lru(folio),
 516                         -folio_nr_pages(folio));
 517         folio_unlock(folio);
 518         folio_putback_lru(folio);
 519 }
 520
 521 static void release_pte_pages(pte_t *pte, pte_t *_pte,
 522                 struct list_head *compound_pagelist)
 523 {
 524         struct folio *folio, *tmp;
 525
 526         while (--_pte >= pte) {
 527                 pte_t pteval = ptep_get(_pte);
 528                 unsigned long pfn;
 529
 530                 if (pte_none(pteval))
 531                         continue;
 532                 pfn = pte_pfn(pteval);
 533                 if (is_zero_pfn(pfn))
 534                         continue;
 535                 folio = pfn_folio(pfn);
 536                 if (folio_test_large(folio))
 537                         continue;
 538                 release_pte_folio(folio);
 539         }
 540
 541         list_for_each_entry_safe(folio, tmp, compound_pagelist, lru) {
 542                 list_del(&folio->lru);
 543                 release_pte_folio(folio);
 544         }
 545 }
 546
 547 static bool is_refcount_suitable(struct folio *folio)
 548 {
 549         int expected_refcount;
 550
 551         expected_refcount = folio_mapcount(folio);
 552         if (folio_test_swapcache(folio))
 553                 expected_refcount += folio_nr_pages(folio);
 554
 555         return folio_ref_count(folio) == expected_refcount;
 556 }
 557
 558 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
 559                                         unsigned long address,
 560                                         pte_t *pte,
 561                                         struct collapse_control *cc,
 562                                         struct list_head *compound_pagelist)
 563 {
 564         struct page *page = NULL;
 565         struct folio *folio = NULL;
 566         pte_t *_pte;
 567         int none_or_zero = 0, shared = 0, result = SCAN_FAIL, referenced = 0;
 568         bool writable = false;
 569
 570         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 571              _pte++, address += PAGE_SIZE) {
 572                 pte_t pteval = ptep_get(_pte);
 573                 if (pte_none(pteval) || (pte_present(pteval) &&
 574                                 is_zero_pfn(pte_pfn(pteval)))) {
 575                         ++none_or_zero;
 576                         if (!userfaultfd_armed(vma) &&
 577                             (!cc->is_khugepaged ||
 578                              none_or_zero <= khugepaged_max_ptes_none)) {
 579                                 continue;
 580                         } else {
 581                                 result = SCAN_EXCEED_NONE_PTE;
 582                                 count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
 583                                 goto out;
 584                         }
 585                 }
 586                 if (!pte_present(pteval)) {
 587                         result = SCAN_PTE_NON_PRESENT;
 588                         goto out;
 589                 }
 590                 if (pte_uffd_wp(pteval)) {
 591                         result = SCAN_PTE_UFFD_WP;
 592                         goto out;
 593                 }
 594                 page = vm_normal_page(vma, address, pteval);
 595                 if (unlikely(!page) || unlikely(is_zone_device_page(page))) {
 596                         result = SCAN_PAGE_NULL;
 597                         goto out;
 598                 }
 599
 600                 folio = page_folio(page);
 601                 VM_BUG_ON_FOLIO(!folio_test_anon(folio), folio);
 602
 603                 /* See hpage_collapse_scan_pmd(). */
 604                 if (folio_likely_mapped_shared(folio)) {
 605                         ++shared;
 606                         if (cc->is_khugepaged &&
 607                             shared > khugepaged_max_ptes_shared) {
 608                                 result = SCAN_EXCEED_SHARED_PTE;
 609                                 count_vm_event(THP_SCAN_EXCEED_SHARED_PTE);
 610                                 goto out;
 611                         }
 612                 }
 613
 614                 if (folio_test_large(folio)) {
 615                         struct folio *f;
 616
 617                         /*
 618                          * Check if we have dealt with the compound page
 619                          * already
 620                          */
 621                         list_for_each_entry(f, compound_pagelist, lru) {
 622                                 if (folio == f)
 623                                         goto next;
 624                         }
 625                 }
 626
 627                 /*
 628                  * We can do it before isolate_lru_page because the
 629                  * page can't be freed from under us. NOTE: PG_lock
 630                  * is needed to serialize against split_huge_page
 631                  * when invoked from the VM.
 632                  */
 633                 if (!folio_trylock(folio)) {
 634                         result = SCAN_PAGE_LOCK;
 635                         goto out;
 636                 }
 637
 638                 /*
 639                  * Check if the page has any GUP (or other external) pins.
 640                  *
 641                  * The page table that maps the page has been already unlinked
 642                  * from the page table tree and this process cannot get
 643                  * an additional pin on the page.
 644                  *
 645                  * New pins can come later if the page is shared across fork,
 646                  * but not from this process. The other process cannot write to
 647                  * the page, only trigger CoW.
 648                  */
 649                 if (!is_refcount_suitable(folio)) {
 650                         folio_unlock(folio);
 651                         result = SCAN_PAGE_COUNT;
 652                         goto out;
 653                 }
 654
 655                 /*
 656                  * Isolate the page to avoid collapsing an hugepage
 657                  * currently in use by the VM.
 658                  */
 659                 if (!folio_isolate_lru(folio)) {
 660                         folio_unlock(folio);
 661                         result = SCAN_DEL_PAGE_LRU;
 662                         goto out;
 663                 }
 664                 node_stat_mod_folio(folio,
 665                                 NR_ISOLATED_ANON + folio_is_file_lru(folio),
 666                                 folio_nr_pages(folio));
 667                 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
 668                 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
 669
 670                 if (folio_test_large(folio))
 671                         list_add_tail(&folio->lru, compound_pagelist);
 672 next:
 673                 /*
 674                  * If collapse was initiated by khugepaged, check that there is
 675                  * enough young pte to justify collapsing the page
 676                  */
 677                 if (cc->is_khugepaged &&
 678                     (pte_young(pteval) || folio_test_young(folio) ||
 679                      folio_test_referenced(folio) || mmu_notifier_test_young(vma->vm_mm,
 680                                                                      address)))
 681                         referenced++;
 682
 683                 if (pte_write(pteval))
 684                         writable = true;
 685         }
 686
 687         if (unlikely(!writable)) {
 688                 result = SCAN_PAGE_RO;
 689         } else if (unlikely(cc->is_khugepaged && !referenced)) {
 690                 result = SCAN_LACK_REFERENCED_PAGE;
 691         } else {
 692                 result = SCAN_SUCCEED;
 693                 trace_mm_collapse_huge_page_isolate(&folio->page, none_or_zero,
 694                                                     referenced, writable, result);
 695                 return result;
 696         }
 697 out:
 698         release_pte_pages(pte, _pte, compound_pagelist);
 699         trace_mm_collapse_huge_page_isolate(&folio->page, none_or_zero,
 700                                             referenced, writable, result);
 701         return result;
 702 }
 703
 704 static void __collapse_huge_page_copy_succeeded(pte_t *pte,
 705                                                 struct vm_area_struct *vma,
 706                                                 unsigned long address,
 707                                                 spinlock_t *ptl,
 708                                                 struct list_head *compound_pagelist)
 709 {
 710         struct folio *src, *tmp;
 711         pte_t *_pte;
 712         pte_t pteval;
 713
 714         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 715              _pte++, address += PAGE_SIZE) {
 716                 pteval = ptep_get(_pte);
 717                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
 718                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
 719                         if (is_zero_pfn(pte_pfn(pteval))) {
 720                                 /*
 721                                  * ptl mostly unnecessary.
 722                                  */
 723                                 spin_lock(ptl);
 724                                 ptep_clear(vma->vm_mm, address, _pte);
 725                                 spin_unlock(ptl);
 726                                 ksm_might_unmap_zero_page(vma->vm_mm, pteval);
 727                         }
 728                 } else {
 729                         struct page *src_page = pte_page(pteval);
 730
 731                         src = page_folio(src_page);
 732                         if (!folio_test_large(src))
 733                                 release_pte_folio(src);
 734                         /*
 735                          * ptl mostly unnecessary, but preempt has to
 736                          * be disabled to update the per-cpu stats
 737                          * inside folio_remove_rmap_pte().
 738                          */
 739                         spin_lock(ptl);
 740                         ptep_clear(vma->vm_mm, address, _pte);
 741                         folio_remove_rmap_pte(src, src_page, vma);
 742                         spin_unlock(ptl);
 743                         free_page_and_swap_cache(src_page);
 744                 }
 745         }
 746
 747         list_for_each_entry_safe(src, tmp, compound_pagelist, lru) {
 748                 list_del(&src->lru);
 749                 node_stat_sub_folio(src, NR_ISOLATED_ANON +
 750                                 folio_is_file_lru(src));
 751                 folio_unlock(src);
 752                 free_swap_cache(src);
 753                 folio_putback_lru(src);
 754         }
 755 }
 756
 757 static void __collapse_huge_page_copy_failed(pte_t *pte,
 758                                              pmd_t *pmd,
 759                                              pmd_t orig_pmd,
 760                                              struct vm_area_struct *vma,
 761                                              struct list_head *compound_pagelist)
 762 {
 763         spinlock_t *pmd_ptl;
 764
 765         /*
 766          * Re-establish the PMD to point to the original page table
 767          * entry. Restoring PMD needs to be done prior to releasing
 768          * pages. Since pages are still isolated and locked here,
 769          * acquiring anon_vma_lock_write is unnecessary.
 770          */
 771         pmd_ptl = pmd_lock(vma->vm_mm, pmd);
 772         pmd_populate(vma->vm_mm, pmd, pmd_pgtable(orig_pmd));
 773         spin_unlock(pmd_ptl);
 774         /*
 775          * Release both raw and compound pages isolated
 776          * in __collapse_huge_page_isolate.
 777          */
 778         release_pte_pages(pte, pte + HPAGE_PMD_NR, compound_pagelist);
 779 }
 780
 781 /*
 782  * __collapse_huge_page_copy - attempts to copy memory contents from raw
 783  * pages to a hugepage. Cleans up the raw pages if copying succeeds;
 784  * otherwise restores the original page table and releases isolated raw pages.
 785  * Returns SCAN_SUCCEED if copying succeeds, otherwise returns SCAN_COPY_MC.
 786  *
 787  * @pte: starting of the PTEs to copy from
 788  * @folio: the new hugepage to copy contents to
 789  * @pmd: pointer to the new hugepage's PMD
 790  * @orig_pmd: the original raw pages' PMD
 791  * @vma: the original raw pages' virtual memory area
 792  * @address: starting address to copy
 793  * @ptl: lock on raw pages' PTEs
 794  * @compound_pagelist: list that stores compound pages
 795  */
 796 static int __collapse_huge_page_copy(pte_t *pte, struct folio *folio,
 797                 pmd_t *pmd, pmd_t orig_pmd, struct vm_area_struct *vma,
 798                 unsigned long address, spinlock_t *ptl,
 799                 struct list_head *compound_pagelist)
 800 {
 801         unsigned int i;
 802         int result = SCAN_SUCCEED;
 803
 804         /*
 805          * Copying pages' contents is subject to memory poison at any iteration.
 806          */
 807         for (i = 0; i < HPAGE_PMD_NR; i++) {
 808                 pte_t pteval = ptep_get(pte + i);
 809                 struct page *page = folio_page(folio, i);
 810                 unsigned long src_addr = address + i * PAGE_SIZE;
 811                 struct page *src_page;
 812
 813                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
 814                         clear_user_highpage(page, src_addr);
 815                         continue;
 816                 }
 817                 src_page = pte_page(pteval);
 818                 if (copy_mc_user_highpage(page, src_page, src_addr, vma) > 0) {
 819                         result = SCAN_COPY_MC;
 820                         break;
 821                 }
 822         }
 823
 824         if (likely(result == SCAN_SUCCEED))
 825                 __collapse_huge_page_copy_succeeded(pte, vma, address, ptl,
 826                                                     compound_pagelist);
 827         else
 828                 __collapse_huge_page_copy_failed(pte, pmd, orig_pmd, vma,
 829                                                  compound_pagelist);
 830
 831         return result;
 832 }
 833
 834 static void khugepaged_alloc_sleep(void)
 835 {
 836         DEFINE_WAIT(wait);
 837
 838         add_wait_queue(&khugepaged_wait, &wait);
 839         __set_current_state(TASK_INTERRUPTIBLE|TASK_FREEZABLE);
 840         schedule_timeout(msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
 841         remove_wait_queue(&khugepaged_wait, &wait);
 842 }
 843
 844 struct collapse_control khugepaged_collapse_control = {
 845         .is_khugepaged = true,
 846 };
 847
 848 static bool hpage_collapse_scan_abort(int nid, struct collapse_control *cc)
 849 {
 850         int i;
 851
 852         /*
 853          * If node_reclaim_mode is disabled, then no extra effort is made to
 854          * allocate memory locally.
 855          */
 856         if (!node_reclaim_enabled())
 857                 return false;
 858
 859         /* If there is a count for this node already, it must be acceptable */
 860         if (cc->node_load[nid])
 861                 return false;
 862
 863         for (i = 0; i < MAX_NUMNODES; i++) {
 864                 if (!cc->node_load[i])
 865                         continue;
 866                 if (node_distance(nid, i) > node_reclaim_distance)
 867                         return true;
 868         }
 869         return false;
 870 }
 871
 872 #define khugepaged_defrag()                                     \
 873         (transparent_hugepage_flags &                           \
 874          (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG))
 875
 876 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
 877 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
 878 {
 879         return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
 880 }
 881
 882 #ifdef CONFIG_NUMA
 883 static int hpage_collapse_find_target_node(struct collapse_control *cc)
 884 {
 885         int nid, target_node = 0, max_value = 0;
 886
 887         /* find first node with max normal pages hit */
 888         for (nid = 0; nid < MAX_NUMNODES; nid++)
 889                 if (cc->node_load[nid] > max_value) {
 890                         max_value = cc->node_load[nid];
 891                         target_node = nid;
 892                 }
 893
 894         for_each_online_node(nid) {
 895                 if (max_value == cc->node_load[nid])
 896                         node_set(nid, cc->alloc_nmask);
 897         }
 898
 899         return target_node;
 900 }
 901 #else
 902 static int hpage_collapse_find_target_node(struct collapse_control *cc)
 903 {
 904         return 0;
 905 }
 906 #endif
 907
 908 /*
 909  * If mmap_lock temporarily dropped, revalidate vma
 910  * before taking mmap_lock.
 911  * Returns enum scan_result value.
 912  */
 913
 914 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
 915                                    bool expect_anon,
 916                                    struct vm_area_struct **vmap,
 917                                    struct collapse_control *cc)
 918 {
 919         struct vm_area_struct *vma;
 920         unsigned long tva_flags = cc->is_khugepaged ? TVA_ENFORCE_SYSFS : 0;
 921
 922         if (unlikely(hpage_collapse_test_exit_or_disable(mm)))
 923                 return SCAN_ANY_PROCESS;
 924
 925         *vmap = vma = find_vma(mm, address);
 926         if (!vma)
 927                 return SCAN_VMA_NULL;
 928
 929         if (!thp_vma_suitable_order(vma, address, PMD_ORDER))
 930                 return SCAN_ADDRESS_RANGE;
 931         if (!thp_vma_allowable_order(vma, vma->vm_flags, tva_flags, PMD_ORDER))
 932                 return SCAN_VMA_CHECK;
 933         /*
 934          * Anon VMA expected, the address may be unmapped then
 935          * remapped to file after khugepaged reaquired the mmap_lock.
 936          *
 937          * thp_vma_allowable_order may return true for qualified file
 938          * vmas.
 939          */
 940         if (expect_anon && (!(*vmap)->anon_vma || !vma_is_anonymous(*vmap)))
 941                 return SCAN_PAGE_ANON;
 942         return SCAN_SUCCEED;
 943 }
 944
 945 static int find_pmd_or_thp_or_none(struct mm_struct *mm,
 946                                    unsigned long address,
 947                                    pmd_t **pmd)
 948 {
 949         pmd_t pmde;
 950
 951         *pmd = mm_find_pmd(mm, address);
 952         if (!*pmd)
 953                 return SCAN_PMD_NULL;
 954
 955         pmde = pmdp_get_lockless(*pmd);
 956         if (pmd_none(pmde))
 957                 return SCAN_PMD_NONE;
 958         if (!pmd_present(pmde))
 959                 return SCAN_PMD_NULL;
 960         if (pmd_trans_huge(pmde))
 961                 return SCAN_PMD_MAPPED;
 962         if (pmd_devmap(pmde))
 963                 return SCAN_PMD_NULL;
 964         if (pmd_bad(pmde))
 965                 return SCAN_PMD_NULL;
 966         return SCAN_SUCCEED;
 967 }
 968
 969 static int check_pmd_still_valid(struct mm_struct *mm,
 970                                  unsigned long address,
 971                                  pmd_t *pmd)
 972 {
 973         pmd_t *new_pmd;
 974         int result = find_pmd_or_thp_or_none(mm, address, &new_pmd);
 975
 976         if (result != SCAN_SUCCEED)
 977                 return result;
 978         if (new_pmd != pmd)
 979                 return SCAN_FAIL;
 980         return SCAN_SUCCEED;
 981 }
 982
 983 /*
 984  * Bring missing pages in from swap, to complete THP collapse.
 985  * Only done if hpage_collapse_scan_pmd believes it is worthwhile.
 986  *
 987  * Called and returns without pte mapped or spinlocks held.
 988  * Returns result: if not SCAN_SUCCEED, mmap_lock has been released.
 989  */
 990 static int __collapse_huge_page_swapin(struct mm_struct *mm,
 991                                        struct vm_area_struct *vma,
 992                                        unsigned long haddr, pmd_t *pmd,
 993                                        int referenced)
 994 {
 995         int swapped_in = 0;
 996         vm_fault_t ret = 0;
 997         unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE);
 998         int result;
 999         pte_t *pte = NULL;
1000         spinlock_t *ptl;
1001
1002         for (address = haddr; address < end; address += PAGE_SIZE) {
1003                 struct vm_fault vmf = {
1004                         .vma = vma,
1005                         .address = address,
1006                         .pgoff = linear_page_index(vma, address),
1007                         .flags = FAULT_FLAG_ALLOW_RETRY,
1008                         .pmd = pmd,
1009                 };
1010
1011                 if (!pte++) {
1012                         pte = pte_offset_map_nolock(mm, pmd, address, &ptl);
1013                         if (!pte) {
1014                                 mmap_read_unlock(mm);
1015                                 result = SCAN_PMD_NULL;
1016                                 goto out;
1017                         }
1018                 }
1019
1020                 vmf.orig_pte = ptep_get_lockless(pte);
1021                 if (!is_swap_pte(vmf.orig_pte))
1022                         continue;
1023
1024                 vmf.pte = pte;
1025                 vmf.ptl = ptl;
1026                 ret = do_swap_page(&vmf);
1027                 /* Which unmaps pte (after perhaps re-checking the entry) */
1028                 pte = NULL;
1029
1030                 /*
1031                  * do_swap_page returns VM_FAULT_RETRY with released mmap_lock.
1032                  * Note we treat VM_FAULT_RETRY as VM_FAULT_ERROR here because
1033                  * we do not retry here and swap entry will remain in pagetable
1034                  * resulting in later failure.
1035                  */
1036                 if (ret & VM_FAULT_RETRY) {
1037                         /* Likely, but not guaranteed, that page lock failed */
1038                         result = SCAN_PAGE_LOCK;
1039                         goto out;
1040                 }
1041                 if (ret & VM_FAULT_ERROR) {
1042                         mmap_read_unlock(mm);
1043                         result = SCAN_FAIL;
1044                         goto out;
1045                 }
1046                 swapped_in++;
1047         }
1048
1049         if (pte)
1050                 pte_unmap(pte);
1051
1052         /* Drain LRU cache to remove extra pin on the swapped in pages */
1053         if (swapped_in)
1054                 lru_add_drain();
1055
1056         result = SCAN_SUCCEED;
1057 out:
1058         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, result);
1059         return result;
1060 }
1061
1062 static int alloc_charge_folio(struct folio **foliop, struct mm_struct *mm,
1063                               struct collapse_control *cc)
1064 {
1065         gfp_t gfp = (cc->is_khugepaged ? alloc_hugepage_khugepaged_gfpmask() :
1066                      GFP_TRANSHUGE);
1067         int node = hpage_collapse_find_target_node(cc);
1068         struct folio *folio;
1069
1070         folio = __folio_alloc(gfp, HPAGE_PMD_ORDER, node, &cc->alloc_nmask);
1071         if (!folio) {
1072                 *foliop = NULL;
1073                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1074                 return SCAN_ALLOC_HUGE_PAGE_FAIL;
1075         }
1076
1077         count_vm_event(THP_COLLAPSE_ALLOC);
1078         if (unlikely(mem_cgroup_charge(folio, mm, gfp))) {
1079                 folio_put(folio);
1080                 *foliop = NULL;
1081                 return SCAN_CGROUP_CHARGE_FAIL;
1082         }
1083
1084         count_memcg_folio_events(folio, THP_COLLAPSE_ALLOC, 1);
1085
1086         *foliop = folio;
1087         return SCAN_SUCCEED;
1088 }
1089
1090 static int collapse_huge_page(struct mm_struct *mm, unsigned long address,
1091                               int referenced, int unmapped,
1092                               struct collapse_control *cc)
1093 {
1094         LIST_HEAD(compound_pagelist);
1095         pmd_t *pmd, _pmd;
1096         pte_t *pte;
1097         pgtable_t pgtable;
1098         struct folio *folio;
1099         spinlock_t *pmd_ptl, *pte_ptl;
1100         int result = SCAN_FAIL;
1101         struct vm_area_struct *vma;
1102         struct mmu_notifier_range range;
1103
1104         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1105
1106         /*
1107          * Before allocating the hugepage, release the mmap_lock read lock.
1108          * The allocation can take potentially a long time if it involves
1109          * sync compaction, and we do not need to hold the mmap_lock during
1110          * that. We will recheck the vma after taking it again in write mode.
1111          */
1112         mmap_read_unlock(mm);
1113
1114         result = alloc_charge_folio(&folio, mm, cc);
1115         if (result != SCAN_SUCCEED)
1116                 goto out_nolock;
1117
1118         mmap_read_lock(mm);
1119         result = hugepage_vma_revalidate(mm, address, true, &vma, cc);
1120         if (result != SCAN_SUCCEED) {
1121                 mmap_read_unlock(mm);
1122                 goto out_nolock;
1123         }
1124
1125         result = find_pmd_or_thp_or_none(mm, address, &pmd);
1126         if (result != SCAN_SUCCEED) {
1127                 mmap_read_unlock(mm);
1128                 goto out_nolock;
1129         }
1130
1131         if (unmapped) {
1132                 /*
1133                  * __collapse_huge_page_swapin will return with mmap_lock
1134                  * released when it fails. So we jump out_nolock directly in
1135                  * that case.  Continuing to collapse causes inconsistency.
1136                  */
1137                 result = __collapse_huge_page_swapin(mm, vma, address, pmd,
1138                                                      referenced);
1139                 if (result != SCAN_SUCCEED)
1140                         goto out_nolock;
1141         }
1142
1143         mmap_read_unlock(mm);
1144         /*
1145          * Prevent all access to pagetables with the exception of
1146          * gup_fast later handled by the ptep_clear_flush and the VM
1147          * handled by the anon_vma lock + PG_lock.
1148          *
1149          * UFFDIO_MOVE is prevented to race as well thanks to the
1150          * mmap_lock.
1151          */
1152         mmap_write_lock(mm);
1153         result = hugepage_vma_revalidate(mm, address, true, &vma, cc);
1154         if (result != SCAN_SUCCEED)
1155                 goto out_up_write;
1156         /* check if the pmd is still valid */
1157         result = check_pmd_still_valid(mm, address, pmd);
1158         if (result != SCAN_SUCCEED)
1159                 goto out_up_write;
1160
1161         vma_start_write(vma);
1162         anon_vma_lock_write(vma->anon_vma);
1163
1164         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, address,
1165                                 address + HPAGE_PMD_SIZE);
1166         mmu_notifier_invalidate_range_start(&range);
1167
1168         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1169         /*
1170          * This removes any huge TLB entry from the CPU so we won't allow
1171          * huge and small TLB entries for the same virtual address to
1172          * avoid the risk of CPU bugs in that area.
1173          *
1174          * Parallel GUP-fast is fine since GUP-fast will back off when
1175          * it detects PMD is changed.
1176          */
1177         _pmd = pmdp_collapse_flush(vma, address, pmd);
1178         spin_unlock(pmd_ptl);
1179         mmu_notifier_invalidate_range_end(&range);
1180         tlb_remove_table_sync_one();
1181
1182         pte = pte_offset_map_lock(mm, &_pmd, address, &pte_ptl);
1183         if (pte) {
1184                 result = __collapse_huge_page_isolate(vma, address, pte, cc,
1185                                                       &compound_pagelist);
1186                 spin_unlock(pte_ptl);
1187         } else {
1188                 result = SCAN_PMD_NULL;
1189         }
1190
1191         if (unlikely(result != SCAN_SUCCEED)) {
1192                 if (pte)
1193                         pte_unmap(pte);
1194                 spin_lock(pmd_ptl);
1195                 BUG_ON(!pmd_none(*pmd));
1196                 /*
1197                  * We can only use set_pmd_at when establishing
1198                  * hugepmds and never for establishing regular pmds that
1199                  * points to regular pagetables. Use pmd_populate for that
1200                  */
1201                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1202                 spin_unlock(pmd_ptl);
1203                 anon_vma_unlock_write(vma->anon_vma);
1204                 goto out_up_write;
1205         }
1206
1207         /*
1208          * All pages are isolated and locked so anon_vma rmap
1209          * can't run anymore.
1210          */
1211         anon_vma_unlock_write(vma->anon_vma);
1212
1213         result = __collapse_huge_page_copy(pte, folio, pmd, _pmd,
1214                                            vma, address, pte_ptl,
1215                                            &compound_pagelist);
1216         pte_unmap(pte);
1217         if (unlikely(result != SCAN_SUCCEED))
1218                 goto out_up_write;
1219
1220         /*
1221          * The smp_wmb() inside __folio_mark_uptodate() ensures the
1222          * copy_huge_page writes become visible before the set_pmd_at()
1223          * write.
1224          */
1225         __folio_mark_uptodate(folio);
1226         pgtable = pmd_pgtable(_pmd);
1227
1228         _pmd = mk_huge_pmd(&folio->page, vma->vm_page_prot);
1229         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1230
1231         spin_lock(pmd_ptl);
1232         BUG_ON(!pmd_none(*pmd));
1233         folio_add_new_anon_rmap(folio, vma, address, RMAP_EXCLUSIVE);
1234         folio_add_lru_vma(folio, vma);
1235         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1236         set_pmd_at(mm, address, pmd, _pmd);
1237         update_mmu_cache_pmd(vma, address, pmd);
1238         spin_unlock(pmd_ptl);
1239
1240         folio = NULL;
1241
1242         result = SCAN_SUCCEED;
1243 out_up_write:
1244         mmap_write_unlock(mm);
1245 out_nolock:
1246         if (folio)
1247                 folio_put(folio);
1248         trace_mm_collapse_huge_page(mm, result == SCAN_SUCCEED, result);
1249         return result;
1250 }
1251
1252 static int hpage_collapse_scan_pmd(struct mm_struct *mm,
1253                                    struct vm_area_struct *vma,
1254                                    unsigned long address, bool *mmap_locked,
1255                                    struct collapse_control *cc)
1256 {
1257         pmd_t *pmd;
1258         pte_t *pte, *_pte;
1259         int result = SCAN_FAIL, referenced = 0;
1260         int none_or_zero = 0, shared = 0;
1261         struct page *page = NULL;
1262         struct folio *folio = NULL;
1263         unsigned long _address;
1264         spinlock_t *ptl;
1265         int node = NUMA_NO_NODE, unmapped = 0;
1266         bool writable = false;
1267
1268         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1269
1270         result = find_pmd_or_thp_or_none(mm, address, &pmd);
1271         if (result != SCAN_SUCCEED)
1272                 goto out;
1273
1274         memset(cc->node_load, 0, sizeof(cc->node_load));
1275         nodes_clear(cc->alloc_nmask);
1276         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1277         if (!pte) {
1278                 result = SCAN_PMD_NULL;
1279                 goto out;
1280         }
1281
1282         for (_address = address, _pte = pte; _pte < pte + HPAGE_PMD_NR;
1283              _pte++, _address += PAGE_SIZE) {
1284                 pte_t pteval = ptep_get(_pte);
1285                 if (is_swap_pte(pteval)) {
1286                         ++unmapped;
1287                         if (!cc->is_khugepaged ||
1288                             unmapped <= khugepaged_max_ptes_swap) {
1289                                 /*
1290                                  * Always be strict with uffd-wp
1291                                  * enabled swap entries.  Please see
1292                                  * comment below for pte_uffd_wp().
1293                                  */
1294                                 if (pte_swp_uffd_wp_any(pteval)) {
1295                                         result = SCAN_PTE_UFFD_WP;
1296                                         goto out_unmap;
1297                                 }
1298                                 continue;
1299                         } else {
1300                                 result = SCAN_EXCEED_SWAP_PTE;
1301                                 count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
1302                                 goto out_unmap;
1303                         }
1304                 }
1305                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1306                         ++none_or_zero;
1307                         if (!userfaultfd_armed(vma) &&
1308                             (!cc->is_khugepaged ||
1309                              none_or_zero <= khugepaged_max_ptes_none)) {
1310                                 continue;
1311                         } else {
1312                                 result = SCAN_EXCEED_NONE_PTE;
1313                                 count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
1314                                 goto out_unmap;
1315                         }
1316                 }
1317                 if (pte_uffd_wp(pteval)) {
1318                         /*
1319                          * Don't collapse the page if any of the small
1320                          * PTEs are armed with uffd write protection.
1321                          * Here we can also mark the new huge pmd as
1322                          * write protected if any of the small ones is
1323                          * marked but that could bring unknown
1324                          * userfault messages that falls outside of
1325                          * the registered range.  So, just be simple.
1326                          */
1327                         result = SCAN_PTE_UFFD_WP;
1328                         goto out_unmap;
1329                 }
1330                 if (pte_write(pteval))
1331                         writable = true;
1332
1333                 page = vm_normal_page(vma, _address, pteval);
1334                 if (unlikely(!page) || unlikely(is_zone_device_page(page))) {
1335                         result = SCAN_PAGE_NULL;
1336                         goto out_unmap;
1337                 }
1338                 folio = page_folio(page);
1339
1340                 if (!folio_test_anon(folio)) {
1341                         result = SCAN_PAGE_ANON;
1342                         goto out_unmap;
1343                 }
1344
1345                 /*
1346                  * We treat a single page as shared if any part of the THP
1347                  * is shared. "False negatives" from
1348                  * folio_likely_mapped_shared() are not expected to matter
1349                  * much in practice.
1350                  */
1351                 if (folio_likely_mapped_shared(folio)) {
1352                         ++shared;
1353                         if (cc->is_khugepaged &&
1354                             shared > khugepaged_max_ptes_shared) {
1355                                 result = SCAN_EXCEED_SHARED_PTE;
1356                                 count_vm_event(THP_SCAN_EXCEED_SHARED_PTE);
1357                                 goto out_unmap;
1358                         }
1359                 }
1360
1361                 /*
1362                  * Record which node the original page is from and save this
1363                  * information to cc->node_load[].
1364                  * Khugepaged will allocate hugepage from the node has the max
1365                  * hit record.
1366                  */
1367                 node = folio_nid(folio);
1368                 if (hpage_collapse_scan_abort(node, cc)) {
1369                         result = SCAN_SCAN_ABORT;
1370                         goto out_unmap;
1371                 }
1372                 cc->node_load[node]++;
1373                 if (!folio_test_lru(folio)) {
1374                         result = SCAN_PAGE_LRU;
1375                         goto out_unmap;
1376                 }
1377                 if (folio_test_locked(folio)) {
1378                         result = SCAN_PAGE_LOCK;
1379                         goto out_unmap;
1380                 }
1381
1382                 /*
1383                  * Check if the page has any GUP (or other external) pins.
1384                  *
1385                  * Here the check may be racy:
1386                  * it may see folio_mapcount() > folio_ref_count().
1387                  * But such case is ephemeral we could always retry collapse
1388                  * later.  However it may report false positive if the page
1389                  * has excessive GUP pins (i.e. 512).  Anyway the same check
1390                  * will be done again later the risk seems low.
1391                  */
1392                 if (!is_refcount_suitable(folio)) {
1393                         result = SCAN_PAGE_COUNT;
1394                         goto out_unmap;
1395                 }
1396
1397                 /*
1398                  * If collapse was initiated by khugepaged, check that there is
1399                  * enough young pte to justify collapsing the page
1400                  */
1401                 if (cc->is_khugepaged &&
1402                     (pte_young(pteval) || folio_test_young(folio) ||
1403                      folio_test_referenced(folio) || mmu_notifier_test_young(vma->vm_mm,
1404                                                                      address)))
1405                         referenced++;
1406         }
1407         if (!writable) {
1408                 result = SCAN_PAGE_RO;
1409         } else if (cc->is_khugepaged &&
1410                    (!referenced ||
1411                     (unmapped && referenced < HPAGE_PMD_NR / 2))) {
1412                 result = SCAN_LACK_REFERENCED_PAGE;
1413         } else {
1414                 result = SCAN_SUCCEED;
1415         }
1416 out_unmap:
1417         pte_unmap_unlock(pte, ptl);
1418         if (result == SCAN_SUCCEED) {
1419                 result = collapse_huge_page(mm, address, referenced,
1420                                             unmapped, cc);
1421                 /* collapse_huge_page will return with the mmap_lock released */
1422                 *mmap_locked = false;
1423         }
1424 out:
1425         trace_mm_khugepaged_scan_pmd(mm, &folio->page, writable, referenced,
1426                                      none_or_zero, result, unmapped);
1427         return result;
1428 }
1429
1430 static void collect_mm_slot(struct khugepaged_mm_slot *mm_slot)
1431 {
1432         struct mm_slot *slot = &mm_slot->slot;
1433         struct mm_struct *mm = slot->mm;
1434
1435         lockdep_assert_held(&khugepaged_mm_lock);
1436
1437         if (hpage_collapse_test_exit(mm)) {
1438                 /* free mm_slot */
1439                 hash_del(&slot->hash);
1440                 list_del(&slot->mm_node);
1441
1442                 /*
1443                  * Not strictly needed because the mm exited already.
1444                  *
1445                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1446                  */
1447
1448                 /* khugepaged_mm_lock actually not necessary for the below */
1449                 mm_slot_free(mm_slot_cache, mm_slot);
1450                 mmdrop(mm);
1451         }
1452 }
1453
1454 #ifdef CONFIG_SHMEM
1455 /* hpage must be locked, and mmap_lock must be held */
1456 static int set_huge_pmd(struct vm_area_struct *vma, unsigned long addr,
1457                         pmd_t *pmdp, struct page *hpage)
1458 {
1459         struct vm_fault vmf = {
1460                 .vma = vma,
1461                 .address = addr,
1462                 .flags = 0,
1463                 .pmd = pmdp,
1464         };
1465
1466         VM_BUG_ON(!PageTransHuge(hpage));
1467         mmap_assert_locked(vma->vm_mm);
1468
1469         if (do_set_pmd(&vmf, hpage))
1470                 return SCAN_FAIL;
1471
1472         get_page(hpage);
1473         return SCAN_SUCCEED;
1474 }
1475
1476 /**
1477  * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at
1478  * address haddr.
1479  *
1480  * @mm: process address space where collapse happens
1481  * @addr: THP collapse address
1482  * @install_pmd: If a huge PMD should be installed
1483  *
1484  * This function checks whether all the PTEs in the PMD are pointing to the
1485  * right THP. If so, retract the page table so the THP can refault in with
1486  * as pmd-mapped. Possibly install a huge PMD mapping the THP.
1487  */
1488 int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
1489                             bool install_pmd)
1490 {
1491         struct mmu_notifier_range range;
1492         bool notified = false;
1493         unsigned long haddr = addr & HPAGE_PMD_MASK;
1494         struct vm_area_struct *vma = vma_lookup(mm, haddr);
1495         struct folio *folio;
1496         pte_t *start_pte, *pte;
1497         pmd_t *pmd, pgt_pmd;
1498         spinlock_t *pml = NULL, *ptl;
1499         int nr_ptes = 0, result = SCAN_FAIL;
1500         int i;
1501
1502         mmap_assert_locked(mm);
1503
1504         /* First check VMA found, in case page tables are being torn down */
1505         if (!vma || !vma->vm_file ||
1506             !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE))
1507                 return SCAN_VMA_CHECK;
1508
1509         /* Fast check before locking page if already PMD-mapped */
1510         result = find_pmd_or_thp_or_none(mm, haddr, &pmd);
1511         if (result == SCAN_PMD_MAPPED)
1512                 return result;
1513
1514         /*
1515          * If we are here, we've succeeded in replacing all the native pages
1516          * in the page cache with a single hugepage. If a mm were to fault-in
1517          * this memory (mapped by a suitably aligned VMA), we'd get the hugepage
1518          * and map it by a PMD, regardless of sysfs THP settings. As such, let's
1519          * analogously elide sysfs THP settings here.
1520          */
1521         if (!thp_vma_allowable_order(vma, vma->vm_flags, 0, PMD_ORDER))
1522                 return SCAN_VMA_CHECK;
1523
1524         /* Keep pmd pgtable for uffd-wp; see comment in retract_page_tables() */
1525         if (userfaultfd_wp(vma))
1526                 return SCAN_PTE_UFFD_WP;
1527
1528         folio = filemap_lock_folio(vma->vm_file->f_mapping,
1529                                linear_page_index(vma, haddr));
1530         if (IS_ERR(folio))
1531                 return SCAN_PAGE_NULL;
1532
1533         if (folio_order(folio) != HPAGE_PMD_ORDER) {
1534                 result = SCAN_PAGE_COMPOUND;
1535                 goto drop_folio;
1536         }
1537
1538         result = find_pmd_or_thp_or_none(mm, haddr, &pmd);
1539         switch (result) {
1540         case SCAN_SUCCEED:
1541                 break;
1542         case SCAN_PMD_NONE:
1543                 /*
1544                  * All pte entries have been removed and pmd cleared.
1545                  * Skip all the pte checks and just update the pmd mapping.
1546                  */
1547                 goto maybe_install_pmd;
1548         default:
1549                 goto drop_folio;
1550         }
1551
1552         result = SCAN_FAIL;
1553         start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1554         if (!start_pte)         /* mmap_lock + page lock should prevent this */
1555                 goto drop_folio;
1556
1557         /* step 1: check all mapped PTEs are to the right huge page */
1558         for (i = 0, addr = haddr, pte = start_pte;
1559              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1560                 struct page *page;
1561                 pte_t ptent = ptep_get(pte);
1562
1563                 /* empty pte, skip */
1564                 if (pte_none(ptent))
1565                         continue;
1566
1567                 /* page swapped out, abort */
1568                 if (!pte_present(ptent)) {
1569                         result = SCAN_PTE_NON_PRESENT;
1570                         goto abort;
1571                 }
1572
1573                 page = vm_normal_page(vma, addr, ptent);
1574                 if (WARN_ON_ONCE(page && is_zone_device_page(page)))
1575                         page = NULL;
1576                 /*
1577                  * Note that uprobe, debugger, or MAP_PRIVATE may change the
1578                  * page table, but the new page will not be a subpage of hpage.
1579                  */
1580                 if (folio_page(folio, i) != page)
1581                         goto abort;
1582         }
1583
1584         pte_unmap_unlock(start_pte, ptl);
1585         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
1586                                 haddr, haddr + HPAGE_PMD_SIZE);
1587         mmu_notifier_invalidate_range_start(&range);
1588         notified = true;
1589
1590         /*
1591          * pmd_lock covers a wider range than ptl, and (if split from mm's
1592          * page_table_lock) ptl nests inside pml. The less time we hold pml,
1593          * the better; but userfaultfd's mfill_atomic_pte() on a private VMA
1594          * inserts a valid as-if-COWed PTE without even looking up page cache.
1595          * So page lock of folio does not protect from it, so we must not drop
1596          * ptl before pgt_pmd is removed, so uffd private needs pml taken now.
1597          */
1598         if (userfaultfd_armed(vma) && !(vma->vm_flags & VM_SHARED))
1599                 pml = pmd_lock(mm, pmd);
1600
1601         start_pte = pte_offset_map_nolock(mm, pmd, haddr, &ptl);
1602         if (!start_pte)         /* mmap_lock + page lock should prevent this */
1603                 goto abort;
1604         if (!pml)
1605                 spin_lock(ptl);
1606         else if (ptl != pml)
1607                 spin_lock_nested(ptl, SINGLE_DEPTH_NESTING);
1608
1609         /* step 2: clear page table and adjust rmap */
1610         for (i = 0, addr = haddr, pte = start_pte;
1611              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1612                 struct page *page;
1613                 pte_t ptent = ptep_get(pte);
1614
1615                 if (pte_none(ptent))
1616                         continue;
1617                 /*
1618                  * We dropped ptl after the first scan, to do the mmu_notifier:
1619                  * page lock stops more PTEs of the folio being faulted in, but
1620                  * does not stop write faults COWing anon copies from existing
1621                  * PTEs; and does not stop those being swapped out or migrated.
1622                  */
1623                 if (!pte_present(ptent)) {
1624                         result = SCAN_PTE_NON_PRESENT;
1625                         goto abort;
1626                 }
1627                 page = vm_normal_page(vma, addr, ptent);
1628                 if (folio_page(folio, i) != page)
1629                         goto abort;
1630
1631                 /*
1632                  * Must clear entry, or a racing truncate may re-remove it.
1633                  * TLB flush can be left until pmdp_collapse_flush() does it.
1634                  * PTE dirty? Shmem page is already dirty; file is read-only.
1635                  */
1636                 ptep_clear(mm, addr, pte);
1637                 folio_remove_rmap_pte(folio, page, vma);
1638                 nr_ptes++;
1639         }
1640
1641         pte_unmap(start_pte);
1642         if (!pml)
1643                 spin_unlock(ptl);
1644
1645         /* step 3: set proper refcount and mm_counters. */
1646         if (nr_ptes) {
1647                 folio_ref_sub(folio, nr_ptes);
1648                 add_mm_counter(mm, mm_counter_file(folio), -nr_ptes);
1649         }
1650
1651         /* step 4: remove empty page table */
1652         if (!pml) {
1653                 pml = pmd_lock(mm, pmd);
1654                 if (ptl != pml)
1655                         spin_lock_nested(ptl, SINGLE_DEPTH_NESTING);
1656         }
1657         pgt_pmd = pmdp_collapse_flush(vma, haddr, pmd);
1658         pmdp_get_lockless_sync();
1659         if (ptl != pml)
1660                 spin_unlock(ptl);
1661         spin_unlock(pml);
1662
1663         mmu_notifier_invalidate_range_end(&range);
1664
1665         mm_dec_nr_ptes(mm);
1666         page_table_check_pte_clear_range(mm, haddr, pgt_pmd);
1667         pte_free_defer(mm, pmd_pgtable(pgt_pmd));
1668
1669 maybe_install_pmd:
1670         /* step 5: install pmd entry */
1671         result = install_pmd
1672                         ? set_huge_pmd(vma, haddr, pmd, &folio->page)
1673                         : SCAN_SUCCEED;
1674         goto drop_folio;
1675 abort:
1676         if (nr_ptes) {
1677                 flush_tlb_mm(mm);
1678                 folio_ref_sub(folio, nr_ptes);
1679                 add_mm_counter(mm, mm_counter_file(folio), -nr_ptes);
1680         }
1681         if (start_pte)
1682                 pte_unmap_unlock(start_pte, ptl);
1683         if (pml && pml != ptl)
1684                 spin_unlock(pml);
1685         if (notified)
1686                 mmu_notifier_invalidate_range_end(&range);
1687 drop_folio:
1688         folio_unlock(folio);
1689         folio_put(folio);
1690         return result;
1691 }
1692
1693 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1694 {
1695         struct vm_area_struct *vma;
1696
1697         i_mmap_lock_read(mapping);
1698         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1699                 struct mmu_notifier_range range;
1700                 struct mm_struct *mm;
1701                 unsigned long addr;
1702                 pmd_t *pmd, pgt_pmd;
1703                 spinlock_t *pml;
1704                 spinlock_t *ptl;
1705                 bool skipped_uffd = false;
1706
1707                 /*
1708                  * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1709                  * got written to. These VMAs are likely not worth removing
1710                  * page tables from, as PMD-mapping is likely to be split later.
1711                  */
1712                 if (READ_ONCE(vma->anon_vma))
1713                         continue;
1714
1715                 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1716                 if (addr & ~HPAGE_PMD_MASK ||
1717                     vma->vm_end < addr + HPAGE_PMD_SIZE)
1718                         continue;
1719
1720                 mm = vma->vm_mm;
1721                 if (find_pmd_or_thp_or_none(mm, addr, &pmd) != SCAN_SUCCEED)
1722                         continue;
1723
1724                 if (hpage_collapse_test_exit(mm))
1725                         continue;
1726                 /*
1727                  * When a vma is registered with uffd-wp, we cannot recycle
1728                  * the page table because there may be pte markers installed.
1729                  * Other vmas can still have the same file mapped hugely, but
1730                  * skip this one: it will always be mapped in small page size
1731                  * for uffd-wp registered ranges.
1732                  */
1733                 if (userfaultfd_wp(vma))
1734                         continue;
1735
1736                 /* PTEs were notified when unmapped; but now for the PMD? */
1737                 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
1738                                         addr, addr + HPAGE_PMD_SIZE);
1739                 mmu_notifier_invalidate_range_start(&range);
1740
1741                 pml = pmd_lock(mm, pmd);
1742                 ptl = pte_lockptr(mm, pmd);
1743                 if (ptl != pml)
1744                         spin_lock_nested(ptl, SINGLE_DEPTH_NESTING);
1745
1746                 /*
1747                  * Huge page lock is still held, so normally the page table
1748                  * must remain empty; and we have already skipped anon_vma
1749                  * and userfaultfd_wp() vmas.  But since the mmap_lock is not
1750                  * held, it is still possible for a racing userfaultfd_ioctl()
1751                  * to have inserted ptes or markers.  Now that we hold ptlock,
1752                  * repeating the anon_vma check protects from one category,
1753                  * and repeating the userfaultfd_wp() check from another.
1754                  */
1755                 if (unlikely(vma->anon_vma || userfaultfd_wp(vma))) {
1756                         skipped_uffd = true;
1757                 } else {
1758                         pgt_pmd = pmdp_collapse_flush(vma, addr, pmd);
1759                         pmdp_get_lockless_sync();
1760                 }
1761
1762                 if (ptl != pml)
1763                         spin_unlock(ptl);
1764                 spin_unlock(pml);
1765
1766                 mmu_notifier_invalidate_range_end(&range);
1767
1768                 if (!skipped_uffd) {
1769                         mm_dec_nr_ptes(mm);
1770                         page_table_check_pte_clear_range(mm, addr, pgt_pmd);
1771                         pte_free_defer(mm, pmd_pgtable(pgt_pmd));
1772                 }
1773         }
1774         i_mmap_unlock_read(mapping);
1775 }
1776
1777 /**
1778  * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1779  *
1780  * @mm: process address space where collapse happens
1781  * @addr: virtual collapse start address
1782  * @file: file that collapse on
1783  * @start: collapse start address
1784  * @cc: collapse context and scratchpad
1785  *
1786  * Basic scheme is simple, details are more complex:
1787  *  - allocate and lock a new huge page;
1788  *  - scan page cache, locking old pages
1789  *    + swap/gup in pages if necessary;
1790  *  - copy data to new page
1791  *  - handle shmem holes
1792  *    + re-validate that holes weren't filled by someone else
1793  *    + check for userfaultfd
1794  *  - finalize updates to the page cache;
1795  *  - if replacing succeeds:
1796  *    + unlock huge page;
1797  *    + free old pages;
1798  *  - if replacing failed;
1799  *    + unlock old pages
1800  *    + unlock and free huge page;
1801  */
1802 static int collapse_file(struct mm_struct *mm, unsigned long addr,
1803                          struct file *file, pgoff_t start,
1804                          struct collapse_control *cc)
1805 {
1806         struct address_space *mapping = file->f_mapping;
1807         struct page *dst;
1808         struct folio *folio, *tmp, *new_folio;
1809         pgoff_t index = 0, end = start + HPAGE_PMD_NR;
1810         LIST_HEAD(pagelist);
1811         XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1812         int nr_none = 0, result = SCAN_SUCCEED;
1813         bool is_shmem = shmem_file(file);
1814
1815         VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1816         VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1817
1818         result = alloc_charge_folio(&new_folio, mm, cc);
1819         if (result != SCAN_SUCCEED)
1820                 goto out;
1821
1822         __folio_set_locked(new_folio);
1823         if (is_shmem)
1824                 __folio_set_swapbacked(new_folio);
1825         new_folio->index = start;
1826         new_folio->mapping = mapping;
1827
1828         /*
1829          * Ensure we have slots for all the pages in the range.  This is
1830          * almost certainly a no-op because most of the pages must be present
1831          */
1832         do {
1833                 xas_lock_irq(&xas);
1834                 xas_create_range(&xas);
1835                 if (!xas_error(&xas))
1836                         break;
1837                 xas_unlock_irq(&xas);
1838                 if (!xas_nomem(&xas, GFP_KERNEL)) {
1839                         result = SCAN_FAIL;
1840                         goto rollback;
1841                 }
1842         } while (1);
1843
1844         for (index = start; index < end; index++) {
1845                 xas_set(&xas, index);
1846                 folio = xas_load(&xas);
1847
1848                 VM_BUG_ON(index != xas.xa_index);
1849                 if (is_shmem) {
1850                         if (!folio) {
1851                                 /*
1852                                  * Stop if extent has been truncated or
1853                                  * hole-punched, and is now completely
1854                                  * empty.
1855                                  */
1856                                 if (index == start) {
1857                                         if (!xas_next_entry(&xas, end - 1)) {
1858                                                 result = SCAN_TRUNCATED;
1859                                                 goto xa_locked;
1860                                         }
1861                                 }
1862                                 nr_none++;
1863                                 continue;
1864                         }
1865
1866                         if (xa_is_value(folio) || !folio_test_uptodate(folio)) {
1867                                 xas_unlock_irq(&xas);
1868                                 /* swap in or instantiate fallocated page */
1869                                 if (shmem_get_folio(mapping->host, index,
1870                                                 &folio, SGP_NOALLOC)) {
1871                                         result = SCAN_FAIL;
1872                                         goto xa_unlocked;
1873                                 }
1874                                 /* drain lru cache to help isolate_lru_page() */
1875                                 lru_add_drain();
1876                         } else if (folio_trylock(folio)) {
1877                                 folio_get(folio);
1878                                 xas_unlock_irq(&xas);
1879                         } else {
1880                                 result = SCAN_PAGE_LOCK;
1881                                 goto xa_locked;
1882                         }
1883                 } else {        /* !is_shmem */
1884                         if (!folio || xa_is_value(folio)) {
1885                                 xas_unlock_irq(&xas);
1886                                 page_cache_sync_readahead(mapping, &file->f_ra,
1887                                                           file, index,
1888                                                           end - index);
1889                                 /* drain lru cache to help isolate_lru_page() */
1890                                 lru_add_drain();
1891                                 folio = filemap_lock_folio(mapping, index);
1892                                 if (IS_ERR(folio)) {
1893                                         result = SCAN_FAIL;
1894                                         goto xa_unlocked;
1895                                 }
1896                         } else if (folio_test_dirty(folio)) {
1897                                 /*
1898                                  * khugepaged only works on read-only fd,
1899                                  * so this page is dirty because it hasn't
1900                                  * been flushed since first write. There
1901                                  * won't be new dirty pages.
1902                                  *
1903                                  * Trigger async flush here and hope the
1904                                  * writeback is done when khugepaged
1905                                  * revisits this page.
1906                                  *
1907                                  * This is a one-off situation. We are not
1908                                  * forcing writeback in loop.
1909                                  */
1910                                 xas_unlock_irq(&xas);
1911                                 filemap_flush(mapping);
1912                                 result = SCAN_FAIL;
1913                                 goto xa_unlocked;
1914                         } else if (folio_test_writeback(folio)) {
1915                                 xas_unlock_irq(&xas);
1916                                 result = SCAN_FAIL;
1917                                 goto xa_unlocked;
1918                         } else if (folio_trylock(folio)) {
1919                                 folio_get(folio);
1920                                 xas_unlock_irq(&xas);
1921                         } else {
1922                                 result = SCAN_PAGE_LOCK;
1923                                 goto xa_locked;
1924                         }
1925                 }
1926
1927                 /*
1928                  * The folio must be locked, so we can drop the i_pages lock
1929                  * without racing with truncate.
1930                  */
1931                 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
1932
1933                 /* make sure the folio is up to date */
1934                 if (unlikely(!folio_test_uptodate(folio))) {
1935                         result = SCAN_FAIL;
1936                         goto out_unlock;
1937                 }
1938
1939                 /*
1940                  * If file was truncated then extended, or hole-punched, before
1941                  * we locked the first folio, then a THP might be there already.
1942                  * This will be discovered on the first iteration.
1943                  */
1944                 if (folio_test_large(folio)) {
1945                         result = folio_order(folio) == HPAGE_PMD_ORDER &&
1946                                         folio->index == start
1947                                         /* Maybe PMD-mapped */
1948                                         ? SCAN_PTE_MAPPED_HUGEPAGE
1949                                         : SCAN_PAGE_COMPOUND;
1950                         goto out_unlock;
1951                 }
1952
1953                 if (folio_mapping(folio) != mapping) {
1954                         result = SCAN_TRUNCATED;
1955                         goto out_unlock;
1956                 }
1957
1958                 if (!is_shmem && (folio_test_dirty(folio) ||
1959                                   folio_test_writeback(folio))) {
1960                         /*
1961                          * khugepaged only works on read-only fd, so this
1962                          * folio is dirty because it hasn't been flushed
1963                          * since first write.
1964                          */
1965                         result = SCAN_FAIL;
1966                         goto out_unlock;
1967                 }
1968
1969                 if (!folio_isolate_lru(folio)) {
1970                         result = SCAN_DEL_PAGE_LRU;
1971                         goto out_unlock;
1972                 }
1973
1974                 if (!filemap_release_folio(folio, GFP_KERNEL)) {
1975                         result = SCAN_PAGE_HAS_PRIVATE;
1976                         folio_putback_lru(folio);
1977                         goto out_unlock;
1978                 }
1979
1980                 if (folio_mapped(folio))
1981                         try_to_unmap(folio,
1982                                         TTU_IGNORE_MLOCK | TTU_BATCH_FLUSH);
1983
1984                 xas_lock_irq(&xas);
1985
1986                 VM_BUG_ON_FOLIO(folio != xa_load(xas.xa, index), folio);
1987
1988                 /*
1989                  * We control three references to the folio:
1990                  *  - we hold a pin on it;
1991                  *  - one reference from page cache;
1992                  *  - one from lru_isolate_folio;
1993                  * If those are the only references, then any new usage
1994                  * of the folio will have to fetch it from the page
1995                  * cache. That requires locking the folio to handle
1996                  * truncate, so any new usage will be blocked until we
1997                  * unlock folio after collapse/during rollback.
1998                  */
1999                 if (folio_ref_count(folio) != 3) {
2000                         result = SCAN_PAGE_COUNT;
2001                         xas_unlock_irq(&xas);
2002                         folio_putback_lru(folio);
2003                         goto out_unlock;
2004                 }
2005
2006                 /*
2007                  * Accumulate the folios that are being collapsed.
2008                  */
2009                 list_add_tail(&folio->lru, &pagelist);
2010                 continue;
2011 out_unlock:
2012                 folio_unlock(folio);
2013                 folio_put(folio);
2014                 goto xa_unlocked;
2015         }
2016
2017         if (!is_shmem) {
2018                 filemap_nr_thps_inc(mapping);
2019                 /*
2020                  * Paired with the fence in do_dentry_open() -> get_write_access()
2021                  * to ensure i_writecount is up to date and the update to nr_thps
2022                  * is visible. Ensures the page cache will be truncated if the
2023                  * file is opened writable.
2024                  */
2025                 smp_mb();
2026                 if (inode_is_open_for_write(mapping->host)) {
2027                         result = SCAN_FAIL;
2028                         filemap_nr_thps_dec(mapping);
2029                 }
2030         }
2031
2032 xa_locked:
2033         xas_unlock_irq(&xas);
2034 xa_unlocked:
2035
2036         /*
2037          * If collapse is successful, flush must be done now before copying.
2038          * If collapse is unsuccessful, does flush actually need to be done?
2039          * Do it anyway, to clear the state.
2040          */
2041         try_to_unmap_flush();
2042
2043         if (result == SCAN_SUCCEED && nr_none &&
2044             !shmem_charge(mapping->host, nr_none))
2045                 result = SCAN_FAIL;
2046         if (result != SCAN_SUCCEED) {
2047                 nr_none = 0;
2048                 goto rollback;
2049         }
2050
2051         /*
2052          * The old folios are locked, so they won't change anymore.
2053          */
2054         index = start;
2055         dst = folio_page(new_folio, 0);
2056         list_for_each_entry(folio, &pagelist, lru) {
2057                 while (index < folio->index) {
2058                         clear_highpage(dst);
2059                         index++;
2060                         dst++;
2061                 }
2062                 if (copy_mc_highpage(dst, folio_page(folio, 0)) > 0) {
2063                         result = SCAN_COPY_MC;
2064                         goto rollback;
2065                 }
2066                 index++;
2067                 dst++;
2068         }
2069         while (index < end) {
2070                 clear_highpage(dst);
2071                 index++;
2072                 dst++;
2073         }
2074
2075         if (nr_none) {
2076                 struct vm_area_struct *vma;
2077                 int nr_none_check = 0;
2078
2079                 i_mmap_lock_read(mapping);
2080                 xas_lock_irq(&xas);
2081
2082                 xas_set(&xas, start);
2083                 for (index = start; index < end; index++) {
2084                         if (!xas_next(&xas)) {
2085                                 xas_store(&xas, XA_RETRY_ENTRY);
2086                                 if (xas_error(&xas)) {
2087                                         result = SCAN_STORE_FAILED;
2088                                         goto immap_locked;
2089                                 }
2090                                 nr_none_check++;
2091                         }
2092                 }
2093
2094                 if (nr_none != nr_none_check) {
2095                         result = SCAN_PAGE_FILLED;
2096                         goto immap_locked;
2097                 }
2098
2099                 /*
2100                  * If userspace observed a missing page in a VMA with
2101                  * a MODE_MISSING userfaultfd, then it might expect a
2102                  * UFFD_EVENT_PAGEFAULT for that page. If so, we need to
2103                  * roll back to avoid suppressing such an event. Since
2104                  * wp/minor userfaultfds don't give userspace any
2105                  * guarantees that the kernel doesn't fill a missing
2106                  * page with a zero page, so they don't matter here.
2107                  *
2108                  * Any userfaultfds registered after this point will
2109                  * not be able to observe any missing pages due to the
2110                  * previously inserted retry entries.
2111                  */
2112                 vma_interval_tree_foreach(vma, &mapping->i_mmap, start, end) {
2113                         if (userfaultfd_missing(vma)) {
2114                                 result = SCAN_EXCEED_NONE_PTE;
2115                                 goto immap_locked;
2116                         }
2117                 }
2118
2119 immap_locked:
2120                 i_mmap_unlock_read(mapping);
2121                 if (result != SCAN_SUCCEED) {
2122                         xas_set(&xas, start);
2123                         for (index = start; index < end; index++) {
2124                                 if (xas_next(&xas) == XA_RETRY_ENTRY)
2125                                         xas_store(&xas, NULL);
2126                         }
2127
2128                         xas_unlock_irq(&xas);
2129                         goto rollback;
2130                 }
2131         } else {
2132                 xas_lock_irq(&xas);
2133         }
2134
2135         if (is_shmem)
2136                 __lruvec_stat_mod_folio(new_folio, NR_SHMEM_THPS, HPAGE_PMD_NR);
2137         else
2138                 __lruvec_stat_mod_folio(new_folio, NR_FILE_THPS, HPAGE_PMD_NR);
2139
2140         if (nr_none) {
2141                 __lruvec_stat_mod_folio(new_folio, NR_FILE_PAGES, nr_none);
2142                 /* nr_none is always 0 for non-shmem. */
2143                 __lruvec_stat_mod_folio(new_folio, NR_SHMEM, nr_none);
2144         }
2145
2146         /*
2147          * Mark new_folio as uptodate before inserting it into the
2148          * page cache so that it isn't mistaken for an fallocated but
2149          * unwritten page.
2150          */
2151         folio_mark_uptodate(new_folio);
2152         folio_ref_add(new_folio, HPAGE_PMD_NR - 1);
2153
2154         if (is_shmem)
2155                 folio_mark_dirty(new_folio);
2156         folio_add_lru(new_folio);
2157
2158         /* Join all the small entries into a single multi-index entry. */
2159         xas_set_order(&xas, start, HPAGE_PMD_ORDER);
2160         xas_store(&xas, new_folio);
2161         WARN_ON_ONCE(xas_error(&xas));
2162         xas_unlock_irq(&xas);
2163
2164         /*
2165          * Remove pte page tables, so we can re-fault the page as huge.
2166          * If MADV_COLLAPSE, adjust result to call collapse_pte_mapped_thp().
2167          */
2168         retract_page_tables(mapping, start);
2169         if (cc && !cc->is_khugepaged)
2170                 result = SCAN_PTE_MAPPED_HUGEPAGE;
2171         folio_unlock(new_folio);
2172
2173         /*
2174          * The collapse has succeeded, so free the old folios.
2175          */
2176         list_for_each_entry_safe(folio, tmp, &pagelist, lru) {
2177                 list_del(&folio->lru);
2178                 folio->mapping = NULL;
2179                 folio_clear_active(folio);
2180                 folio_clear_unevictable(folio);
2181                 folio_unlock(folio);
2182                 folio_put_refs(folio, 3);
2183         }
2184
2185         goto out;
2186
2187 rollback:
2188         /* Something went wrong: roll back page cache changes */
2189         if (nr_none) {
2190                 xas_lock_irq(&xas);
2191                 mapping->nrpages -= nr_none;
2192                 xas_unlock_irq(&xas);
2193                 shmem_uncharge(mapping->host, nr_none);
2194         }
2195
2196         list_for_each_entry_safe(folio, tmp, &pagelist, lru) {
2197                 list_del(&folio->lru);
2198                 folio_unlock(folio);
2199                 folio_putback_lru(folio);
2200                 folio_put(folio);
2201         }
2202         /*
2203          * Undo the updates of filemap_nr_thps_inc for non-SHMEM
2204          * file only. This undo is not needed unless failure is
2205          * due to SCAN_COPY_MC.
2206          */
2207         if (!is_shmem && result == SCAN_COPY_MC) {
2208                 filemap_nr_thps_dec(mapping);
2209                 /*
2210                  * Paired with the fence in do_dentry_open() -> get_write_access()
2211                  * to ensure the update to nr_thps is visible.
2212                  */
2213                 smp_mb();
2214         }
2215
2216         new_folio->mapping = NULL;
2217
2218         folio_unlock(new_folio);
2219         folio_put(new_folio);
2220 out:
2221         VM_BUG_ON(!list_empty(&pagelist));
2222         trace_mm_khugepaged_collapse_file(mm, new_folio, index, is_shmem, addr, file, HPAGE_PMD_NR, result);
2223         return result;
2224 }
2225
2226 static int hpage_collapse_scan_file(struct mm_struct *mm, unsigned long addr,
2227                                     struct file *file, pgoff_t start,
2228                                     struct collapse_control *cc)
2229 {
2230         struct folio *folio = NULL;
2231         struct address_space *mapping = file->f_mapping;
2232         XA_STATE(xas, &mapping->i_pages, start);
2233         int present, swap;
2234         int node = NUMA_NO_NODE;
2235         int result = SCAN_SUCCEED;
2236
2237         present = 0;
2238         swap = 0;
2239         memset(cc->node_load, 0, sizeof(cc->node_load));
2240         nodes_clear(cc->alloc_nmask);
2241         rcu_read_lock();
2242         xas_for_each(&xas, folio, start + HPAGE_PMD_NR - 1) {
2243                 if (xas_retry(&xas, folio))
2244                         continue;
2245
2246                 if (xa_is_value(folio)) {
2247                         ++swap;
2248                         if (cc->is_khugepaged &&
2249                             swap > khugepaged_max_ptes_swap) {
2250                                 result = SCAN_EXCEED_SWAP_PTE;
2251                                 count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
2252                                 break;
2253                         }
2254                         continue;
2255                 }
2256
2257                 /*
2258                  * TODO: khugepaged should compact smaller compound pages
2259                  * into a PMD sized page
2260                  */
2261                 if (folio_test_large(folio)) {
2262                         result = folio_order(folio) == HPAGE_PMD_ORDER &&
2263                                         folio->index == start
2264                                         /* Maybe PMD-mapped */
2265                                         ? SCAN_PTE_MAPPED_HUGEPAGE
2266                                         : SCAN_PAGE_COMPOUND;
2267                         /*
2268                          * For SCAN_PTE_MAPPED_HUGEPAGE, further processing
2269                          * by the caller won't touch the page cache, and so
2270                          * it's safe to skip LRU and refcount checks before
2271                          * returning.
2272                          */
2273                         break;
2274                 }
2275
2276                 node = folio_nid(folio);
2277                 if (hpage_collapse_scan_abort(node, cc)) {
2278                         result = SCAN_SCAN_ABORT;
2279                         break;
2280                 }
2281                 cc->node_load[node]++;
2282
2283                 if (!folio_test_lru(folio)) {
2284                         result = SCAN_PAGE_LRU;
2285                         break;
2286                 }
2287
2288                 if (folio_ref_count(folio) !=
2289                     1 + folio_mapcount(folio) + folio_test_private(folio)) {
2290                         result = SCAN_PAGE_COUNT;
2291                         break;
2292                 }
2293
2294                 /*
2295                  * We probably should check if the folio is referenced
2296                  * here, but nobody would transfer pte_young() to
2297                  * folio_test_referenced() for us.  And rmap walk here
2298                  * is just too costly...
2299                  */
2300
2301                 present++;
2302
2303                 if (need_resched()) {
2304                         xas_pause(&xas);
2305                         cond_resched_rcu();
2306                 }
2307         }
2308         rcu_read_unlock();
2309
2310         if (result == SCAN_SUCCEED) {
2311                 if (cc->is_khugepaged &&
2312                     present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2313                         result = SCAN_EXCEED_NONE_PTE;
2314                         count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
2315                 } else {
2316                         result = collapse_file(mm, addr, file, start, cc);
2317                 }
2318         }
2319
2320         trace_mm_khugepaged_scan_file(mm, folio, file, present, swap, result);
2321         return result;
2322 }
2323 #else
2324 static int hpage_collapse_scan_file(struct mm_struct *mm, unsigned long addr,
2325                                     struct file *file, pgoff_t start,
2326                                     struct collapse_control *cc)
2327 {
2328         BUILD_BUG();
2329 }
2330 #endif
2331
2332 static unsigned int khugepaged_scan_mm_slot(unsigned int pages, int *result,
2333                                             struct collapse_control *cc)
2334         __releases(&khugepaged_mm_lock)
2335         __acquires(&khugepaged_mm_lock)
2336 {
2337         struct vma_iterator vmi;
2338         struct khugepaged_mm_slot *mm_slot;
2339         struct mm_slot *slot;
2340         struct mm_struct *mm;
2341         struct vm_area_struct *vma;
2342         int progress = 0;
2343
2344         VM_BUG_ON(!pages);
2345         lockdep_assert_held(&khugepaged_mm_lock);
2346         *result = SCAN_FAIL;
2347
2348         if (khugepaged_scan.mm_slot) {
2349                 mm_slot = khugepaged_scan.mm_slot;
2350                 slot = &mm_slot->slot;
2351         } else {
2352                 slot = list_entry(khugepaged_scan.mm_head.next,
2353                                      struct mm_slot, mm_node);
2354                 mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot);
2355                 khugepaged_scan.address = 0;
2356                 khugepaged_scan.mm_slot = mm_slot;
2357         }
2358         spin_unlock(&khugepaged_mm_lock);
2359
2360         mm = slot->mm;
2361         /*
2362          * Don't wait for semaphore (to avoid long wait times).  Just move to
2363          * the next mm on the list.
2364          */
2365         vma = NULL;
2366         if (unlikely(!mmap_read_trylock(mm)))
2367                 goto breakouterloop_mmap_lock;
2368
2369         progress++;
2370         if (unlikely(hpage_collapse_test_exit_or_disable(mm)))
2371                 goto breakouterloop;
2372
2373         vma_iter_init(&vmi, mm, khugepaged_scan.address);
2374         for_each_vma(vmi, vma) {
2375                 unsigned long hstart, hend;
2376
2377                 cond_resched();
2378                 if (unlikely(hpage_collapse_test_exit_or_disable(mm))) {
2379                         progress++;
2380                         break;
2381                 }
2382                 if (!thp_vma_allowable_order(vma, vma->vm_flags,
2383                                         TVA_ENFORCE_SYSFS, PMD_ORDER)) {
2384 skip:
2385                         progress++;
2386                         continue;
2387                 }
2388                 hstart = round_up(vma->vm_start, HPAGE_PMD_SIZE);
2389                 hend = round_down(vma->vm_end, HPAGE_PMD_SIZE);
2390                 if (khugepaged_scan.address > hend)
2391                         goto skip;
2392                 if (khugepaged_scan.address < hstart)
2393                         khugepaged_scan.address = hstart;
2394                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2395
2396                 while (khugepaged_scan.address < hend) {
2397                         bool mmap_locked = true;
2398
2399                         cond_resched();
2400                         if (unlikely(hpage_collapse_test_exit_or_disable(mm)))
2401                                 goto breakouterloop;
2402
2403                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2404                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2405                                   hend);
2406                         if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2407                                 struct file *file = get_file(vma->vm_file);
2408                                 pgoff_t pgoff = linear_page_index(vma,
2409                                                 khugepaged_scan.address);
2410
2411                                 mmap_read_unlock(mm);
2412                                 mmap_locked = false;
2413                                 *result = hpage_collapse_scan_file(mm,
2414                                         khugepaged_scan.address, file, pgoff, cc);
2415                                 fput(file);
2416                                 if (*result == SCAN_PTE_MAPPED_HUGEPAGE) {
2417                                         mmap_read_lock(mm);
2418                                         if (hpage_collapse_test_exit_or_disable(mm))
2419                                                 goto breakouterloop;
2420                                         *result = collapse_pte_mapped_thp(mm,
2421                                                 khugepaged_scan.address, false);
2422                                         if (*result == SCAN_PMD_MAPPED)
2423                                                 *result = SCAN_SUCCEED;
2424                                         mmap_read_unlock(mm);
2425                                 }
2426                         } else {
2427                                 *result = hpage_collapse_scan_pmd(mm, vma,
2428                                         khugepaged_scan.address, &mmap_locked, cc);
2429                         }
2430
2431                         if (*result == SCAN_SUCCEED)
2432                                 ++khugepaged_pages_collapsed;
2433
2434                         /* move to next address */
2435                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2436                         progress += HPAGE_PMD_NR;
2437                         if (!mmap_locked)
2438                                 /*
2439                                  * We released mmap_lock so break loop.  Note
2440                                  * that we drop mmap_lock before all hugepage
2441                                  * allocations, so if allocation fails, we are
2442                                  * guaranteed to break here and report the
2443                                  * correct result back to caller.
2444                                  */
2445                                 goto breakouterloop_mmap_lock;
2446                         if (progress >= pages)
2447                                 goto breakouterloop;
2448                 }
2449         }
2450 breakouterloop:
2451         mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2452 breakouterloop_mmap_lock:
2453
2454         spin_lock(&khugepaged_mm_lock);
2455         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2456         /*
2457          * Release the current mm_slot if this mm is about to die, or
2458          * if we scanned all vmas of this mm.
2459          */
2460         if (hpage_collapse_test_exit(mm) || !vma) {
2461                 /*
2462                  * Make sure that if mm_users is reaching zero while
2463                  * khugepaged runs here, khugepaged_exit will find
2464                  * mm_slot not pointing to the exiting mm.
2465                  */
2466                 if (slot->mm_node.next != &khugepaged_scan.mm_head) {
2467                         slot = list_entry(slot->mm_node.next,
2468                                           struct mm_slot, mm_node);
2469                         khugepaged_scan.mm_slot =
2470                                 mm_slot_entry(slot, struct khugepaged_mm_slot, slot);
2471                         khugepaged_scan.address = 0;
2472                 } else {
2473                         khugepaged_scan.mm_slot = NULL;
2474                         khugepaged_full_scans++;
2475                 }
2476
2477                 collect_mm_slot(mm_slot);
2478         }
2479
2480         return progress;
2481 }
2482
2483 static int khugepaged_has_work(void)
2484 {
2485         return !list_empty(&khugepaged_scan.mm_head) && hugepage_pmd_enabled();
2486 }
2487
2488 static int khugepaged_wait_event(void)
2489 {
2490         return !list_empty(&khugepaged_scan.mm_head) ||
2491                 kthread_should_stop();
2492 }
2493
2494 static void khugepaged_do_scan(struct collapse_control *cc)
2495 {
2496         unsigned int progress = 0, pass_through_head = 0;
2497         unsigned int pages = READ_ONCE(khugepaged_pages_to_scan);
2498         bool wait = true;
2499         int result = SCAN_SUCCEED;
2500
2501         lru_add_drain_all();
2502
2503         while (true) {
2504                 cond_resched();
2505
2506                 if (unlikely(kthread_should_stop()))
2507                         break;
2508
2509                 spin_lock(&khugepaged_mm_lock);
2510                 if (!khugepaged_scan.mm_slot)
2511                         pass_through_head++;
2512                 if (khugepaged_has_work() &&
2513                     pass_through_head < 2)
2514                         progress += khugepaged_scan_mm_slot(pages - progress,
2515                                                             &result, cc);
2516                 else
2517                         progress = pages;
2518                 spin_unlock(&khugepaged_mm_lock);
2519
2520                 if (progress >= pages)
2521                         break;
2522
2523                 if (result == SCAN_ALLOC_HUGE_PAGE_FAIL) {
2524                         /*
2525                          * If fail to allocate the first time, try to sleep for
2526                          * a while.  When hit again, cancel the scan.
2527                          */
2528                         if (!wait)
2529                                 break;
2530                         wait = false;
2531                         khugepaged_alloc_sleep();
2532                 }
2533         }
2534 }
2535
2536 static bool khugepaged_should_wakeup(void)
2537 {
2538         return kthread_should_stop() ||
2539                time_after_eq(jiffies, khugepaged_sleep_expire);
2540 }
2541
2542 static void khugepaged_wait_work(void)
2543 {
2544         if (khugepaged_has_work()) {
2545                 const unsigned long scan_sleep_jiffies =
2546                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2547
2548                 if (!scan_sleep_jiffies)
2549                         return;
2550
2551                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2552                 wait_event_freezable_timeout(khugepaged_wait,
2553                                              khugepaged_should_wakeup(),
2554                                              scan_sleep_jiffies);
2555                 return;
2556         }
2557
2558         if (hugepage_pmd_enabled())
2559                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2560 }
2561
2562 static int khugepaged(void *none)
2563 {
2564         struct khugepaged_mm_slot *mm_slot;
2565
2566         set_freezable();
2567         set_user_nice(current, MAX_NICE);
2568
2569         while (!kthread_should_stop()) {
2570                 khugepaged_do_scan(&khugepaged_collapse_control);
2571                 khugepaged_wait_work();
2572         }
2573
2574         spin_lock(&khugepaged_mm_lock);
2575         mm_slot = khugepaged_scan.mm_slot;
2576         khugepaged_scan.mm_slot = NULL;
2577         if (mm_slot)
2578                 collect_mm_slot(mm_slot);
2579         spin_unlock(&khugepaged_mm_lock);
2580         return 0;
2581 }
2582
2583 static void set_recommended_min_free_kbytes(void)
2584 {
2585         struct zone *zone;
2586         int nr_zones = 0;
2587         unsigned long recommended_min;
2588
2589         if (!hugepage_pmd_enabled()) {
2590                 calculate_min_free_kbytes();
2591                 goto update_wmarks;
2592         }
2593
2594         for_each_populated_zone(zone) {
2595                 /*
2596                  * We don't need to worry about fragmentation of
2597                  * ZONE_MOVABLE since it only has movable pages.
2598                  */
2599                 if (zone_idx(zone) > gfp_zone(GFP_USER))
2600                         continue;
2601
2602                 nr_zones++;
2603         }
2604
2605         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2606         recommended_min = pageblock_nr_pages * nr_zones * 2;
2607
2608         /*
2609          * Make sure that on average at least two pageblocks are almost free
2610          * of another type, one for a migratetype to fall back to and a
2611          * second to avoid subsequent fallbacks of other types There are 3
2612          * MIGRATE_TYPES we care about.
2613          */
2614         recommended_min += pageblock_nr_pages * nr_zones *
2615                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2616
2617         /* don't ever allow to reserve more than 5% of the lowmem */
2618         recommended_min = min(recommended_min,
2619                               (unsigned long) nr_free_buffer_pages() / 20);
2620         recommended_min <<= (PAGE_SHIFT-10);
2621
2622         if (recommended_min > min_free_kbytes) {
2623                 if (user_min_free_kbytes >= 0)
2624                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2625                                 min_free_kbytes, recommended_min);
2626
2627                 min_free_kbytes = recommended_min;
2628         }
2629
2630 update_wmarks:
2631         setup_per_zone_wmarks();
2632 }
2633
2634 int start_stop_khugepaged(void)
2635 {
2636         int err = 0;
2637
2638         mutex_lock(&khugepaged_mutex);
2639         if (hugepage_pmd_enabled()) {
2640                 if (!khugepaged_thread)
2641                         khugepaged_thread = kthread_run(khugepaged, NULL,
2642                                                         "khugepaged");
2643                 if (IS_ERR(khugepaged_thread)) {
2644                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2645                         err = PTR_ERR(khugepaged_thread);
2646                         khugepaged_thread = NULL;
2647                         goto fail;
2648                 }
2649
2650                 if (!list_empty(&khugepaged_scan.mm_head))
2651                         wake_up_interruptible(&khugepaged_wait);
2652         } else if (khugepaged_thread) {
2653                 kthread_stop(khugepaged_thread);
2654                 khugepaged_thread = NULL;
2655         }
2656         set_recommended_min_free_kbytes();
2657 fail:
2658         mutex_unlock(&khugepaged_mutex);
2659         return err;
2660 }
2661
2662 void khugepaged_min_free_kbytes_update(void)
2663 {
2664         mutex_lock(&khugepaged_mutex);
2665         if (hugepage_pmd_enabled() && khugepaged_thread)
2666                 set_recommended_min_free_kbytes();
2667         mutex_unlock(&khugepaged_mutex);
2668 }
2669
2670 bool current_is_khugepaged(void)
2671 {
2672         return kthread_func(current) == khugepaged;
2673 }
2674
2675 static int madvise_collapse_errno(enum scan_result r)
2676 {
2677         /*
2678          * MADV_COLLAPSE breaks from existing madvise(2) conventions to provide
2679          * actionable feedback to caller, so they may take an appropriate
2680          * fallback measure depending on the nature of the failure.
2681          */
2682         switch (r) {
2683         case SCAN_ALLOC_HUGE_PAGE_FAIL:
2684                 return -ENOMEM;
2685         case SCAN_CGROUP_CHARGE_FAIL:
2686         case SCAN_EXCEED_NONE_PTE:
2687                 return -EBUSY;
2688         /* Resource temporary unavailable - trying again might succeed */
2689         case SCAN_PAGE_COUNT:
2690         case SCAN_PAGE_LOCK:
2691         case SCAN_PAGE_LRU:
2692         case SCAN_DEL_PAGE_LRU:
2693         case SCAN_PAGE_FILLED:
2694                 return -EAGAIN;
2695         /*
2696          * Other: Trying again likely not to succeed / error intrinsic to
2697          * specified memory range. khugepaged likely won't be able to collapse
2698          * either.
2699          */
2700         default:
2701                 return -EINVAL;
2702         }
2703 }
2704
2705 int madvise_collapse(struct vm_area_struct *vma, struct vm_area_struct **prev,
2706                      unsigned long start, unsigned long end)
2707 {
2708         struct collapse_control *cc;
2709         struct mm_struct *mm = vma->vm_mm;
2710         unsigned long hstart, hend, addr;
2711         int thps = 0, last_fail = SCAN_FAIL;
2712         bool mmap_locked = true;
2713
2714         BUG_ON(vma->vm_start > start);
2715         BUG_ON(vma->vm_end < end);
2716
2717         *prev = vma;
2718
2719         if (!thp_vma_allowable_order(vma, vma->vm_flags, 0, PMD_ORDER))
2720                 return -EINVAL;
2721
2722         cc = kmalloc(sizeof(*cc), GFP_KERNEL);
2723         if (!cc)
2724                 return -ENOMEM;
2725         cc->is_khugepaged = false;
2726
2727         mmgrab(mm);
2728         lru_add_drain_all();
2729
2730         hstart = (start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2731         hend = end & HPAGE_PMD_MASK;
2732
2733         for (addr = hstart; addr < hend; addr += HPAGE_PMD_SIZE) {
2734                 int result = SCAN_FAIL;
2735
2736                 if (!mmap_locked) {
2737                         cond_resched();
2738                         mmap_read_lock(mm);
2739                         mmap_locked = true;
2740                         result = hugepage_vma_revalidate(mm, addr, false, &vma,
2741                                                          cc);
2742                         if (result  != SCAN_SUCCEED) {
2743                                 last_fail = result;
2744                                 goto out_nolock;
2745                         }
2746
2747                         hend = min(hend, vma->vm_end & HPAGE_PMD_MASK);
2748                 }
2749                 mmap_assert_locked(mm);
2750                 memset(cc->node_load, 0, sizeof(cc->node_load));
2751                 nodes_clear(cc->alloc_nmask);
2752                 if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2753                         struct file *file = get_file(vma->vm_file);
2754                         pgoff_t pgoff = linear_page_index(vma, addr);
2755
2756                         mmap_read_unlock(mm);
2757                         mmap_locked = false;
2758                         result = hpage_collapse_scan_file(mm, addr, file, pgoff,
2759                                                           cc);
2760                         fput(file);
2761                 } else {
2762                         result = hpage_collapse_scan_pmd(mm, vma, addr,
2763                                                          &mmap_locked, cc);
2764                 }
2765                 if (!mmap_locked)
2766                         *prev = NULL;  /* Tell caller we dropped mmap_lock */
2767
2768 handle_result:
2769                 switch (result) {
2770                 case SCAN_SUCCEED:
2771                 case SCAN_PMD_MAPPED:
2772                         ++thps;
2773                         break;
2774                 case SCAN_PTE_MAPPED_HUGEPAGE:
2775                         BUG_ON(mmap_locked);
2776                         BUG_ON(*prev);
2777                         mmap_read_lock(mm);
2778                         result = collapse_pte_mapped_thp(mm, addr, true);
2779                         mmap_read_unlock(mm);
2780                         goto handle_result;
2781                 /* Whitelisted set of results where continuing OK */
2782                 case SCAN_PMD_NULL:
2783                 case SCAN_PTE_NON_PRESENT:
2784                 case SCAN_PTE_UFFD_WP:
2785                 case SCAN_PAGE_RO:
2786                 case SCAN_LACK_REFERENCED_PAGE:
2787                 case SCAN_PAGE_NULL:
2788                 case SCAN_PAGE_COUNT:
2789                 case SCAN_PAGE_LOCK:
2790                 case SCAN_PAGE_COMPOUND:
2791                 case SCAN_PAGE_LRU:
2792                 case SCAN_DEL_PAGE_LRU:
2793                         last_fail = result;
2794                         break;
2795                 default:
2796                         last_fail = result;
2797                         /* Other error, exit */
2798                         goto out_maybelock;
2799                 }
2800         }
2801
2802 out_maybelock:
2803         /* Caller expects us to hold mmap_lock on return */
2804         if (!mmap_locked)
2805                 mmap_read_lock(mm);
2806 out_nolock:
2807         mmap_assert_locked(mm);
2808         mmdrop(mm);
2809         kfree(cc);
2810
2811         return thps == ((hend - hstart) >> HPAGE_PMD_SHIFT) ? 0
2812                         : madvise_collapse_errno(last_fail);
2813 }