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