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