1 // SPDX-License-Identifier: GPL-2.0 2 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 3 4 #include <linux/mm.h> 5 #include <linux/sched.h> 6 #include <linux/sched/mm.h> 7 #include <linux/sched/coredump.h> 8 #include <linux/mmu_notifier.h> 9 #include <linux/rmap.h> 10 #include <linux/swap.h> 11 #include <linux/mm_inline.h> 12 #include <linux/kthread.h> 13 #include <linux/khugepaged.h> 14 #include <linux/freezer.h> 15 #include <linux/mman.h> 16 #include <linux/hashtable.h> 17 #include <linux/userfaultfd_k.h> 18 #include <linux/page_idle.h> 19 #include <linux/swapops.h> 20 #include <linux/shmem_fs.h> 21 22 #include <asm/tlb.h> 23 #include <asm/pgalloc.h> 24 #include "internal.h" 25 26 enum scan_result { 27 SCAN_FAIL, 28 SCAN_SUCCEED, 29 SCAN_PMD_NULL, 30 SCAN_EXCEED_NONE_PTE, 31 SCAN_PTE_NON_PRESENT, 32 SCAN_PAGE_RO, 33 SCAN_LACK_REFERENCED_PAGE, 34 SCAN_PAGE_NULL, 35 SCAN_SCAN_ABORT, 36 SCAN_PAGE_COUNT, 37 SCAN_PAGE_LRU, 38 SCAN_PAGE_LOCK, 39 SCAN_PAGE_ANON, 40 SCAN_PAGE_COMPOUND, 41 SCAN_ANY_PROCESS, 42 SCAN_VMA_NULL, 43 SCAN_VMA_CHECK, 44 SCAN_ADDRESS_RANGE, 45 SCAN_SWAP_CACHE_PAGE, 46 SCAN_DEL_PAGE_LRU, 47 SCAN_ALLOC_HUGE_PAGE_FAIL, 48 SCAN_CGROUP_CHARGE_FAIL, 49 SCAN_EXCEED_SWAP_PTE, 50 SCAN_TRUNCATED, 51 SCAN_PAGE_HAS_PRIVATE, 52 }; 53 54 #define CREATE_TRACE_POINTS 55 #include <trace/events/huge_memory.h> 56 57 /* default scan 8*512 pte (or vmas) every 30 second */ 58 static unsigned int khugepaged_pages_to_scan __read_mostly; 59 static unsigned int khugepaged_pages_collapsed; 60 static unsigned int khugepaged_full_scans; 61 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000; 62 /* during fragmentation poll the hugepage allocator once every minute */ 63 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000; 64 static unsigned long khugepaged_sleep_expire; 65 static DEFINE_SPINLOCK(khugepaged_mm_lock); 66 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait); 67 /* 68 * default collapse hugepages if there is at least one pte mapped like 69 * it would have happened if the vma was large enough during page 70 * fault. 71 */ 72 static unsigned int khugepaged_max_ptes_none __read_mostly; 73 static unsigned int khugepaged_max_ptes_swap __read_mostly; 74 75 #define MM_SLOTS_HASH_BITS 10 76 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); 77 78 static struct kmem_cache *mm_slot_cache __read_mostly; 79 80 #define MAX_PTE_MAPPED_THP 8 81 82 /** 83 * struct mm_slot - hash lookup from mm to mm_slot 84 * @hash: hash collision list 85 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head 86 * @mm: the mm that this information is valid for 87 */ 88 struct mm_slot { 89 struct hlist_node hash; 90 struct list_head mm_node; 91 struct mm_struct *mm; 92 93 /* pte-mapped THP in this mm */ 94 int nr_pte_mapped_thp; 95 unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP]; 96 }; 97 98 /** 99 * struct khugepaged_scan - cursor for scanning 100 * @mm_head: the head of the mm list to scan 101 * @mm_slot: the current mm_slot we are scanning 102 * @address: the next address inside that to be scanned 103 * 104 * There is only the one khugepaged_scan instance of this cursor structure. 105 */ 106 struct khugepaged_scan { 107 struct list_head mm_head; 108 struct mm_slot *mm_slot; 109 unsigned long address; 110 }; 111 112 static struct khugepaged_scan khugepaged_scan = { 113 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head), 114 }; 115 116 #ifdef CONFIG_SYSFS 117 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj, 118 struct kobj_attribute *attr, 119 char *buf) 120 { 121 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs); 122 } 123 124 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj, 125 struct kobj_attribute *attr, 126 const char *buf, size_t count) 127 { 128 unsigned long msecs; 129 int err; 130 131 err = kstrtoul(buf, 10, &msecs); 132 if (err || msecs > UINT_MAX) 133 return -EINVAL; 134 135 khugepaged_scan_sleep_millisecs = msecs; 136 khugepaged_sleep_expire = 0; 137 wake_up_interruptible(&khugepaged_wait); 138 139 return count; 140 } 141 static struct kobj_attribute scan_sleep_millisecs_attr = 142 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show, 143 scan_sleep_millisecs_store); 144 145 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj, 146 struct kobj_attribute *attr, 147 char *buf) 148 { 149 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs); 150 } 151 152 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj, 153 struct kobj_attribute *attr, 154 const char *buf, size_t count) 155 { 156 unsigned long msecs; 157 int err; 158 159 err = kstrtoul(buf, 10, &msecs); 160 if (err || msecs > UINT_MAX) 161 return -EINVAL; 162 163 khugepaged_alloc_sleep_millisecs = msecs; 164 khugepaged_sleep_expire = 0; 165 wake_up_interruptible(&khugepaged_wait); 166 167 return count; 168 } 169 static struct kobj_attribute alloc_sleep_millisecs_attr = 170 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show, 171 alloc_sleep_millisecs_store); 172 173 static ssize_t pages_to_scan_show(struct kobject *kobj, 174 struct kobj_attribute *attr, 175 char *buf) 176 { 177 return sprintf(buf, "%u\n", khugepaged_pages_to_scan); 178 } 179 static ssize_t pages_to_scan_store(struct kobject *kobj, 180 struct kobj_attribute *attr, 181 const char *buf, size_t count) 182 { 183 int err; 184 unsigned long pages; 185 186 err = kstrtoul(buf, 10, &pages); 187 if (err || !pages || pages > UINT_MAX) 188 return -EINVAL; 189 190 khugepaged_pages_to_scan = pages; 191 192 return count; 193 } 194 static struct kobj_attribute pages_to_scan_attr = 195 __ATTR(pages_to_scan, 0644, pages_to_scan_show, 196 pages_to_scan_store); 197 198 static ssize_t pages_collapsed_show(struct kobject *kobj, 199 struct kobj_attribute *attr, 200 char *buf) 201 { 202 return sprintf(buf, "%u\n", khugepaged_pages_collapsed); 203 } 204 static struct kobj_attribute pages_collapsed_attr = 205 __ATTR_RO(pages_collapsed); 206 207 static ssize_t full_scans_show(struct kobject *kobj, 208 struct kobj_attribute *attr, 209 char *buf) 210 { 211 return sprintf(buf, "%u\n", khugepaged_full_scans); 212 } 213 static struct kobj_attribute full_scans_attr = 214 __ATTR_RO(full_scans); 215 216 static ssize_t khugepaged_defrag_show(struct kobject *kobj, 217 struct kobj_attribute *attr, char *buf) 218 { 219 return single_hugepage_flag_show(kobj, attr, buf, 220 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); 221 } 222 static ssize_t khugepaged_defrag_store(struct kobject *kobj, 223 struct kobj_attribute *attr, 224 const char *buf, size_t count) 225 { 226 return single_hugepage_flag_store(kobj, attr, buf, count, 227 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); 228 } 229 static struct kobj_attribute khugepaged_defrag_attr = 230 __ATTR(defrag, 0644, khugepaged_defrag_show, 231 khugepaged_defrag_store); 232 233 /* 234 * max_ptes_none controls if khugepaged should collapse hugepages over 235 * any unmapped ptes in turn potentially increasing the memory 236 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not 237 * reduce the available free memory in the system as it 238 * runs. Increasing max_ptes_none will instead potentially reduce the 239 * free memory in the system during the khugepaged scan. 240 */ 241 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj, 242 struct kobj_attribute *attr, 243 char *buf) 244 { 245 return sprintf(buf, "%u\n", khugepaged_max_ptes_none); 246 } 247 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj, 248 struct kobj_attribute *attr, 249 const char *buf, size_t count) 250 { 251 int err; 252 unsigned long max_ptes_none; 253 254 err = kstrtoul(buf, 10, &max_ptes_none); 255 if (err || max_ptes_none > HPAGE_PMD_NR-1) 256 return -EINVAL; 257 258 khugepaged_max_ptes_none = max_ptes_none; 259 260 return count; 261 } 262 static struct kobj_attribute khugepaged_max_ptes_none_attr = 263 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show, 264 khugepaged_max_ptes_none_store); 265 266 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj, 267 struct kobj_attribute *attr, 268 char *buf) 269 { 270 return sprintf(buf, "%u\n", khugepaged_max_ptes_swap); 271 } 272 273 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj, 274 struct kobj_attribute *attr, 275 const char *buf, size_t count) 276 { 277 int err; 278 unsigned long max_ptes_swap; 279 280 err = kstrtoul(buf, 10, &max_ptes_swap); 281 if (err || max_ptes_swap > HPAGE_PMD_NR-1) 282 return -EINVAL; 283 284 khugepaged_max_ptes_swap = max_ptes_swap; 285 286 return count; 287 } 288 289 static struct kobj_attribute khugepaged_max_ptes_swap_attr = 290 __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show, 291 khugepaged_max_ptes_swap_store); 292 293 static struct attribute *khugepaged_attr[] = { 294 &khugepaged_defrag_attr.attr, 295 &khugepaged_max_ptes_none_attr.attr, 296 &pages_to_scan_attr.attr, 297 &pages_collapsed_attr.attr, 298 &full_scans_attr.attr, 299 &scan_sleep_millisecs_attr.attr, 300 &alloc_sleep_millisecs_attr.attr, 301 &khugepaged_max_ptes_swap_attr.attr, 302 NULL, 303 }; 304 305 struct attribute_group khugepaged_attr_group = { 306 .attrs = khugepaged_attr, 307 .name = "khugepaged", 308 }; 309 #endif /* CONFIG_SYSFS */ 310 311 #define VM_NO_KHUGEPAGED (VM_SPECIAL | VM_HUGETLB) 312 313 int hugepage_madvise(struct vm_area_struct *vma, 314 unsigned long *vm_flags, int advice) 315 { 316 switch (advice) { 317 case MADV_HUGEPAGE: 318 #ifdef CONFIG_S390 319 /* 320 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390 321 * can't handle this properly after s390_enable_sie, so we simply 322 * ignore the madvise to prevent qemu from causing a SIGSEGV. 323 */ 324 if (mm_has_pgste(vma->vm_mm)) 325 return 0; 326 #endif 327 *vm_flags &= ~VM_NOHUGEPAGE; 328 *vm_flags |= VM_HUGEPAGE; 329 /* 330 * If the vma become good for khugepaged to scan, 331 * register it here without waiting a page fault that 332 * may not happen any time soon. 333 */ 334 if (!(*vm_flags & VM_NO_KHUGEPAGED) && 335 khugepaged_enter_vma_merge(vma, *vm_flags)) 336 return -ENOMEM; 337 break; 338 case MADV_NOHUGEPAGE: 339 *vm_flags &= ~VM_HUGEPAGE; 340 *vm_flags |= VM_NOHUGEPAGE; 341 /* 342 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning 343 * this vma even if we leave the mm registered in khugepaged if 344 * it got registered before VM_NOHUGEPAGE was set. 345 */ 346 break; 347 } 348 349 return 0; 350 } 351 352 int __init khugepaged_init(void) 353 { 354 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot", 355 sizeof(struct mm_slot), 356 __alignof__(struct mm_slot), 0, NULL); 357 if (!mm_slot_cache) 358 return -ENOMEM; 359 360 khugepaged_pages_to_scan = HPAGE_PMD_NR * 8; 361 khugepaged_max_ptes_none = HPAGE_PMD_NR - 1; 362 khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8; 363 364 return 0; 365 } 366 367 void __init khugepaged_destroy(void) 368 { 369 kmem_cache_destroy(mm_slot_cache); 370 } 371 372 static inline struct mm_slot *alloc_mm_slot(void) 373 { 374 if (!mm_slot_cache) /* initialization failed */ 375 return NULL; 376 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL); 377 } 378 379 static inline void free_mm_slot(struct mm_slot *mm_slot) 380 { 381 kmem_cache_free(mm_slot_cache, mm_slot); 382 } 383 384 static struct mm_slot *get_mm_slot(struct mm_struct *mm) 385 { 386 struct mm_slot *mm_slot; 387 388 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm) 389 if (mm == mm_slot->mm) 390 return mm_slot; 391 392 return NULL; 393 } 394 395 static void insert_to_mm_slots_hash(struct mm_struct *mm, 396 struct mm_slot *mm_slot) 397 { 398 mm_slot->mm = mm; 399 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm); 400 } 401 402 static inline int khugepaged_test_exit(struct mm_struct *mm) 403 { 404 return atomic_read(&mm->mm_users) == 0; 405 } 406 407 static bool hugepage_vma_check(struct vm_area_struct *vma, 408 unsigned long vm_flags) 409 { 410 if ((!(vm_flags & VM_HUGEPAGE) && !khugepaged_always()) || 411 (vm_flags & VM_NOHUGEPAGE) || 412 test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags)) 413 return false; 414 415 if (shmem_file(vma->vm_file) || 416 (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && 417 vma->vm_file && 418 (vm_flags & VM_DENYWRITE))) { 419 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) 420 return false; 421 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff, 422 HPAGE_PMD_NR); 423 } 424 if (!vma->anon_vma || vma->vm_ops) 425 return false; 426 if (is_vma_temporary_stack(vma)) 427 return false; 428 return !(vm_flags & VM_NO_KHUGEPAGED); 429 } 430 431 int __khugepaged_enter(struct mm_struct *mm) 432 { 433 struct mm_slot *mm_slot; 434 int wakeup; 435 436 mm_slot = alloc_mm_slot(); 437 if (!mm_slot) 438 return -ENOMEM; 439 440 /* __khugepaged_exit() must not run from under us */ 441 VM_BUG_ON_MM(khugepaged_test_exit(mm), mm); 442 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) { 443 free_mm_slot(mm_slot); 444 return 0; 445 } 446 447 spin_lock(&khugepaged_mm_lock); 448 insert_to_mm_slots_hash(mm, mm_slot); 449 /* 450 * Insert just behind the scanning cursor, to let the area settle 451 * down a little. 452 */ 453 wakeup = list_empty(&khugepaged_scan.mm_head); 454 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head); 455 spin_unlock(&khugepaged_mm_lock); 456 457 mmgrab(mm); 458 if (wakeup) 459 wake_up_interruptible(&khugepaged_wait); 460 461 return 0; 462 } 463 464 int khugepaged_enter_vma_merge(struct vm_area_struct *vma, 465 unsigned long vm_flags) 466 { 467 unsigned long hstart, hend; 468 469 /* 470 * khugepaged only supports read-only files for non-shmem files. 471 * khugepaged does not yet work on special mappings. And 472 * file-private shmem THP is not supported. 473 */ 474 if (!hugepage_vma_check(vma, vm_flags)) 475 return 0; 476 477 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; 478 hend = vma->vm_end & HPAGE_PMD_MASK; 479 if (hstart < hend) 480 return khugepaged_enter(vma, vm_flags); 481 return 0; 482 } 483 484 void __khugepaged_exit(struct mm_struct *mm) 485 { 486 struct mm_slot *mm_slot; 487 int free = 0; 488 489 spin_lock(&khugepaged_mm_lock); 490 mm_slot = get_mm_slot(mm); 491 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) { 492 hash_del(&mm_slot->hash); 493 list_del(&mm_slot->mm_node); 494 free = 1; 495 } 496 spin_unlock(&khugepaged_mm_lock); 497 498 if (free) { 499 clear_bit(MMF_VM_HUGEPAGE, &mm->flags); 500 free_mm_slot(mm_slot); 501 mmdrop(mm); 502 } else if (mm_slot) { 503 /* 504 * This is required to serialize against 505 * khugepaged_test_exit() (which is guaranteed to run 506 * under mmap sem read mode). Stop here (after we 507 * return all pagetables will be destroyed) until 508 * khugepaged has finished working on the pagetables 509 * under the mmap_sem. 510 */ 511 down_write(&mm->mmap_sem); 512 up_write(&mm->mmap_sem); 513 } 514 } 515 516 static void release_pte_page(struct page *page) 517 { 518 dec_node_page_state(page, NR_ISOLATED_ANON + page_is_file_cache(page)); 519 unlock_page(page); 520 putback_lru_page(page); 521 } 522 523 static void release_pte_pages(pte_t *pte, pte_t *_pte) 524 { 525 while (--_pte >= pte) { 526 pte_t pteval = *_pte; 527 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval))) 528 release_pte_page(pte_page(pteval)); 529 } 530 } 531 532 static int __collapse_huge_page_isolate(struct vm_area_struct *vma, 533 unsigned long address, 534 pte_t *pte) 535 { 536 struct page *page = NULL; 537 pte_t *_pte; 538 int none_or_zero = 0, result = 0, referenced = 0; 539 bool writable = false; 540 541 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; 542 _pte++, address += PAGE_SIZE) { 543 pte_t pteval = *_pte; 544 if (pte_none(pteval) || (pte_present(pteval) && 545 is_zero_pfn(pte_pfn(pteval)))) { 546 if (!userfaultfd_armed(vma) && 547 ++none_or_zero <= khugepaged_max_ptes_none) { 548 continue; 549 } else { 550 result = SCAN_EXCEED_NONE_PTE; 551 goto out; 552 } 553 } 554 if (!pte_present(pteval)) { 555 result = SCAN_PTE_NON_PRESENT; 556 goto out; 557 } 558 page = vm_normal_page(vma, address, pteval); 559 if (unlikely(!page)) { 560 result = SCAN_PAGE_NULL; 561 goto out; 562 } 563 564 /* TODO: teach khugepaged to collapse THP mapped with pte */ 565 if (PageCompound(page)) { 566 result = SCAN_PAGE_COMPOUND; 567 goto out; 568 } 569 570 VM_BUG_ON_PAGE(!PageAnon(page), page); 571 572 /* 573 * We can do it before isolate_lru_page because the 574 * page can't be freed from under us. NOTE: PG_lock 575 * is needed to serialize against split_huge_page 576 * when invoked from the VM. 577 */ 578 if (!trylock_page(page)) { 579 result = SCAN_PAGE_LOCK; 580 goto out; 581 } 582 583 /* 584 * cannot use mapcount: can't collapse if there's a gup pin. 585 * The page must only be referenced by the scanned process 586 * and page swap cache. 587 */ 588 if (page_count(page) != 1 + PageSwapCache(page)) { 589 unlock_page(page); 590 result = SCAN_PAGE_COUNT; 591 goto out; 592 } 593 if (pte_write(pteval)) { 594 writable = true; 595 } else { 596 if (PageSwapCache(page) && 597 !reuse_swap_page(page, NULL)) { 598 unlock_page(page); 599 result = SCAN_SWAP_CACHE_PAGE; 600 goto out; 601 } 602 /* 603 * Page is not in the swap cache. It can be collapsed 604 * into a THP. 605 */ 606 } 607 608 /* 609 * Isolate the page to avoid collapsing an hugepage 610 * currently in use by the VM. 611 */ 612 if (isolate_lru_page(page)) { 613 unlock_page(page); 614 result = SCAN_DEL_PAGE_LRU; 615 goto out; 616 } 617 inc_node_page_state(page, 618 NR_ISOLATED_ANON + page_is_file_cache(page)); 619 VM_BUG_ON_PAGE(!PageLocked(page), page); 620 VM_BUG_ON_PAGE(PageLRU(page), page); 621 622 /* There should be enough young pte to collapse the page */ 623 if (pte_young(pteval) || 624 page_is_young(page) || PageReferenced(page) || 625 mmu_notifier_test_young(vma->vm_mm, address)) 626 referenced++; 627 } 628 if (likely(writable)) { 629 if (likely(referenced)) { 630 result = SCAN_SUCCEED; 631 trace_mm_collapse_huge_page_isolate(page, none_or_zero, 632 referenced, writable, result); 633 return 1; 634 } 635 } else { 636 result = SCAN_PAGE_RO; 637 } 638 639 out: 640 release_pte_pages(pte, _pte); 641 trace_mm_collapse_huge_page_isolate(page, none_or_zero, 642 referenced, writable, result); 643 return 0; 644 } 645 646 static void __collapse_huge_page_copy(pte_t *pte, struct page *page, 647 struct vm_area_struct *vma, 648 unsigned long address, 649 spinlock_t *ptl) 650 { 651 pte_t *_pte; 652 for (_pte = pte; _pte < pte + HPAGE_PMD_NR; 653 _pte++, page++, address += PAGE_SIZE) { 654 pte_t pteval = *_pte; 655 struct page *src_page; 656 657 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { 658 clear_user_highpage(page, address); 659 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); 660 if (is_zero_pfn(pte_pfn(pteval))) { 661 /* 662 * ptl mostly unnecessary. 663 */ 664 spin_lock(ptl); 665 /* 666 * paravirt calls inside pte_clear here are 667 * superfluous. 668 */ 669 pte_clear(vma->vm_mm, address, _pte); 670 spin_unlock(ptl); 671 } 672 } else { 673 src_page = pte_page(pteval); 674 copy_user_highpage(page, src_page, address, vma); 675 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page); 676 release_pte_page(src_page); 677 /* 678 * ptl mostly unnecessary, but preempt has to 679 * be disabled to update the per-cpu stats 680 * inside page_remove_rmap(). 681 */ 682 spin_lock(ptl); 683 /* 684 * paravirt calls inside pte_clear here are 685 * superfluous. 686 */ 687 pte_clear(vma->vm_mm, address, _pte); 688 page_remove_rmap(src_page, false); 689 spin_unlock(ptl); 690 free_page_and_swap_cache(src_page); 691 } 692 } 693 } 694 695 static void khugepaged_alloc_sleep(void) 696 { 697 DEFINE_WAIT(wait); 698 699 add_wait_queue(&khugepaged_wait, &wait); 700 freezable_schedule_timeout_interruptible( 701 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs)); 702 remove_wait_queue(&khugepaged_wait, &wait); 703 } 704 705 static int khugepaged_node_load[MAX_NUMNODES]; 706 707 static bool khugepaged_scan_abort(int nid) 708 { 709 int i; 710 711 /* 712 * If node_reclaim_mode is disabled, then no extra effort is made to 713 * allocate memory locally. 714 */ 715 if (!node_reclaim_mode) 716 return false; 717 718 /* If there is a count for this node already, it must be acceptable */ 719 if (khugepaged_node_load[nid]) 720 return false; 721 722 for (i = 0; i < MAX_NUMNODES; i++) { 723 if (!khugepaged_node_load[i]) 724 continue; 725 if (node_distance(nid, i) > node_reclaim_distance) 726 return true; 727 } 728 return false; 729 } 730 731 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */ 732 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void) 733 { 734 return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT; 735 } 736 737 #ifdef CONFIG_NUMA 738 static int khugepaged_find_target_node(void) 739 { 740 static int last_khugepaged_target_node = NUMA_NO_NODE; 741 int nid, target_node = 0, max_value = 0; 742 743 /* find first node with max normal pages hit */ 744 for (nid = 0; nid < MAX_NUMNODES; nid++) 745 if (khugepaged_node_load[nid] > max_value) { 746 max_value = khugepaged_node_load[nid]; 747 target_node = nid; 748 } 749 750 /* do some balance if several nodes have the same hit record */ 751 if (target_node <= last_khugepaged_target_node) 752 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES; 753 nid++) 754 if (max_value == khugepaged_node_load[nid]) { 755 target_node = nid; 756 break; 757 } 758 759 last_khugepaged_target_node = target_node; 760 return target_node; 761 } 762 763 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) 764 { 765 if (IS_ERR(*hpage)) { 766 if (!*wait) 767 return false; 768 769 *wait = false; 770 *hpage = NULL; 771 khugepaged_alloc_sleep(); 772 } else if (*hpage) { 773 put_page(*hpage); 774 *hpage = NULL; 775 } 776 777 return true; 778 } 779 780 static struct page * 781 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node) 782 { 783 VM_BUG_ON_PAGE(*hpage, *hpage); 784 785 *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER); 786 if (unlikely(!*hpage)) { 787 count_vm_event(THP_COLLAPSE_ALLOC_FAILED); 788 *hpage = ERR_PTR(-ENOMEM); 789 return NULL; 790 } 791 792 prep_transhuge_page(*hpage); 793 count_vm_event(THP_COLLAPSE_ALLOC); 794 return *hpage; 795 } 796 #else 797 static int khugepaged_find_target_node(void) 798 { 799 return 0; 800 } 801 802 static inline struct page *alloc_khugepaged_hugepage(void) 803 { 804 struct page *page; 805 806 page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(), 807 HPAGE_PMD_ORDER); 808 if (page) 809 prep_transhuge_page(page); 810 return page; 811 } 812 813 static struct page *khugepaged_alloc_hugepage(bool *wait) 814 { 815 struct page *hpage; 816 817 do { 818 hpage = alloc_khugepaged_hugepage(); 819 if (!hpage) { 820 count_vm_event(THP_COLLAPSE_ALLOC_FAILED); 821 if (!*wait) 822 return NULL; 823 824 *wait = false; 825 khugepaged_alloc_sleep(); 826 } else 827 count_vm_event(THP_COLLAPSE_ALLOC); 828 } while (unlikely(!hpage) && likely(khugepaged_enabled())); 829 830 return hpage; 831 } 832 833 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) 834 { 835 if (!*hpage) 836 *hpage = khugepaged_alloc_hugepage(wait); 837 838 if (unlikely(!*hpage)) 839 return false; 840 841 return true; 842 } 843 844 static struct page * 845 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node) 846 { 847 VM_BUG_ON(!*hpage); 848 849 return *hpage; 850 } 851 #endif 852 853 /* 854 * If mmap_sem temporarily dropped, revalidate vma 855 * before taking mmap_sem. 856 * Return 0 if succeeds, otherwise return none-zero 857 * value (scan code). 858 */ 859 860 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address, 861 struct vm_area_struct **vmap) 862 { 863 struct vm_area_struct *vma; 864 unsigned long hstart, hend; 865 866 if (unlikely(khugepaged_test_exit(mm))) 867 return SCAN_ANY_PROCESS; 868 869 *vmap = vma = find_vma(mm, address); 870 if (!vma) 871 return SCAN_VMA_NULL; 872 873 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; 874 hend = vma->vm_end & HPAGE_PMD_MASK; 875 if (address < hstart || address + HPAGE_PMD_SIZE > hend) 876 return SCAN_ADDRESS_RANGE; 877 if (!hugepage_vma_check(vma, vma->vm_flags)) 878 return SCAN_VMA_CHECK; 879 return 0; 880 } 881 882 /* 883 * Bring missing pages in from swap, to complete THP collapse. 884 * Only done if khugepaged_scan_pmd believes it is worthwhile. 885 * 886 * Called and returns without pte mapped or spinlocks held, 887 * but with mmap_sem held to protect against vma changes. 888 */ 889 890 static bool __collapse_huge_page_swapin(struct mm_struct *mm, 891 struct vm_area_struct *vma, 892 unsigned long address, pmd_t *pmd, 893 int referenced) 894 { 895 int swapped_in = 0; 896 vm_fault_t ret = 0; 897 struct vm_fault vmf = { 898 .vma = vma, 899 .address = address, 900 .flags = FAULT_FLAG_ALLOW_RETRY, 901 .pmd = pmd, 902 .pgoff = linear_page_index(vma, address), 903 }; 904 905 /* we only decide to swapin, if there is enough young ptes */ 906 if (referenced < HPAGE_PMD_NR/2) { 907 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0); 908 return false; 909 } 910 vmf.pte = pte_offset_map(pmd, address); 911 for (; vmf.address < address + HPAGE_PMD_NR*PAGE_SIZE; 912 vmf.pte++, vmf.address += PAGE_SIZE) { 913 vmf.orig_pte = *vmf.pte; 914 if (!is_swap_pte(vmf.orig_pte)) 915 continue; 916 swapped_in++; 917 ret = do_swap_page(&vmf); 918 919 /* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */ 920 if (ret & VM_FAULT_RETRY) { 921 down_read(&mm->mmap_sem); 922 if (hugepage_vma_revalidate(mm, address, &vmf.vma)) { 923 /* vma is no longer available, don't continue to swapin */ 924 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0); 925 return false; 926 } 927 /* check if the pmd is still valid */ 928 if (mm_find_pmd(mm, address) != pmd) { 929 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0); 930 return false; 931 } 932 } 933 if (ret & VM_FAULT_ERROR) { 934 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0); 935 return false; 936 } 937 /* pte is unmapped now, we need to map it */ 938 vmf.pte = pte_offset_map(pmd, vmf.address); 939 } 940 vmf.pte--; 941 pte_unmap(vmf.pte); 942 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1); 943 return true; 944 } 945 946 static void collapse_huge_page(struct mm_struct *mm, 947 unsigned long address, 948 struct page **hpage, 949 int node, int referenced) 950 { 951 pmd_t *pmd, _pmd; 952 pte_t *pte; 953 pgtable_t pgtable; 954 struct page *new_page; 955 spinlock_t *pmd_ptl, *pte_ptl; 956 int isolated = 0, result = 0; 957 struct mem_cgroup *memcg; 958 struct vm_area_struct *vma; 959 struct mmu_notifier_range range; 960 gfp_t gfp; 961 962 VM_BUG_ON(address & ~HPAGE_PMD_MASK); 963 964 /* Only allocate from the target node */ 965 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE; 966 967 /* 968 * Before allocating the hugepage, release the mmap_sem read lock. 969 * The allocation can take potentially a long time if it involves 970 * sync compaction, and we do not need to hold the mmap_sem during 971 * that. We will recheck the vma after taking it again in write mode. 972 */ 973 up_read(&mm->mmap_sem); 974 new_page = khugepaged_alloc_page(hpage, gfp, node); 975 if (!new_page) { 976 result = SCAN_ALLOC_HUGE_PAGE_FAIL; 977 goto out_nolock; 978 } 979 980 if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) { 981 result = SCAN_CGROUP_CHARGE_FAIL; 982 goto out_nolock; 983 } 984 985 down_read(&mm->mmap_sem); 986 result = hugepage_vma_revalidate(mm, address, &vma); 987 if (result) { 988 mem_cgroup_cancel_charge(new_page, memcg, true); 989 up_read(&mm->mmap_sem); 990 goto out_nolock; 991 } 992 993 pmd = mm_find_pmd(mm, address); 994 if (!pmd) { 995 result = SCAN_PMD_NULL; 996 mem_cgroup_cancel_charge(new_page, memcg, true); 997 up_read(&mm->mmap_sem); 998 goto out_nolock; 999 } 1000 1001 /* 1002 * __collapse_huge_page_swapin always returns with mmap_sem locked. 1003 * If it fails, we release mmap_sem and jump out_nolock. 1004 * Continuing to collapse causes inconsistency. 1005 */ 1006 if (!__collapse_huge_page_swapin(mm, vma, address, pmd, referenced)) { 1007 mem_cgroup_cancel_charge(new_page, memcg, true); 1008 up_read(&mm->mmap_sem); 1009 goto out_nolock; 1010 } 1011 1012 up_read(&mm->mmap_sem); 1013 /* 1014 * Prevent all access to pagetables with the exception of 1015 * gup_fast later handled by the ptep_clear_flush and the VM 1016 * handled by the anon_vma lock + PG_lock. 1017 */ 1018 down_write(&mm->mmap_sem); 1019 result = SCAN_ANY_PROCESS; 1020 if (!mmget_still_valid(mm)) 1021 goto out; 1022 result = hugepage_vma_revalidate(mm, address, &vma); 1023 if (result) 1024 goto out; 1025 /* check if the pmd is still valid */ 1026 if (mm_find_pmd(mm, address) != pmd) 1027 goto out; 1028 1029 anon_vma_lock_write(vma->anon_vma); 1030 1031 pte = pte_offset_map(pmd, address); 1032 pte_ptl = pte_lockptr(mm, pmd); 1033 1034 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm, 1035 address, address + HPAGE_PMD_SIZE); 1036 mmu_notifier_invalidate_range_start(&range); 1037 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */ 1038 /* 1039 * After this gup_fast can't run anymore. This also removes 1040 * any huge TLB entry from the CPU so we won't allow 1041 * huge and small TLB entries for the same virtual address 1042 * to avoid the risk of CPU bugs in that area. 1043 */ 1044 _pmd = pmdp_collapse_flush(vma, address, pmd); 1045 spin_unlock(pmd_ptl); 1046 mmu_notifier_invalidate_range_end(&range); 1047 1048 spin_lock(pte_ptl); 1049 isolated = __collapse_huge_page_isolate(vma, address, pte); 1050 spin_unlock(pte_ptl); 1051 1052 if (unlikely(!isolated)) { 1053 pte_unmap(pte); 1054 spin_lock(pmd_ptl); 1055 BUG_ON(!pmd_none(*pmd)); 1056 /* 1057 * We can only use set_pmd_at when establishing 1058 * hugepmds and never for establishing regular pmds that 1059 * points to regular pagetables. Use pmd_populate for that 1060 */ 1061 pmd_populate(mm, pmd, pmd_pgtable(_pmd)); 1062 spin_unlock(pmd_ptl); 1063 anon_vma_unlock_write(vma->anon_vma); 1064 result = SCAN_FAIL; 1065 goto out; 1066 } 1067 1068 /* 1069 * All pages are isolated and locked so anon_vma rmap 1070 * can't run anymore. 1071 */ 1072 anon_vma_unlock_write(vma->anon_vma); 1073 1074 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl); 1075 pte_unmap(pte); 1076 __SetPageUptodate(new_page); 1077 pgtable = pmd_pgtable(_pmd); 1078 1079 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot); 1080 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma); 1081 1082 /* 1083 * spin_lock() below is not the equivalent of smp_wmb(), so 1084 * this is needed to avoid the copy_huge_page writes to become 1085 * visible after the set_pmd_at() write. 1086 */ 1087 smp_wmb(); 1088 1089 spin_lock(pmd_ptl); 1090 BUG_ON(!pmd_none(*pmd)); 1091 page_add_new_anon_rmap(new_page, vma, address, true); 1092 mem_cgroup_commit_charge(new_page, memcg, false, true); 1093 count_memcg_events(memcg, THP_COLLAPSE_ALLOC, 1); 1094 lru_cache_add_active_or_unevictable(new_page, vma); 1095 pgtable_trans_huge_deposit(mm, pmd, pgtable); 1096 set_pmd_at(mm, address, pmd, _pmd); 1097 update_mmu_cache_pmd(vma, address, pmd); 1098 spin_unlock(pmd_ptl); 1099 1100 *hpage = NULL; 1101 1102 khugepaged_pages_collapsed++; 1103 result = SCAN_SUCCEED; 1104 out_up_write: 1105 up_write(&mm->mmap_sem); 1106 out_nolock: 1107 trace_mm_collapse_huge_page(mm, isolated, result); 1108 return; 1109 out: 1110 mem_cgroup_cancel_charge(new_page, memcg, true); 1111 goto out_up_write; 1112 } 1113 1114 static int khugepaged_scan_pmd(struct mm_struct *mm, 1115 struct vm_area_struct *vma, 1116 unsigned long address, 1117 struct page **hpage) 1118 { 1119 pmd_t *pmd; 1120 pte_t *pte, *_pte; 1121 int ret = 0, none_or_zero = 0, result = 0, referenced = 0; 1122 struct page *page = NULL; 1123 unsigned long _address; 1124 spinlock_t *ptl; 1125 int node = NUMA_NO_NODE, unmapped = 0; 1126 bool writable = false; 1127 1128 VM_BUG_ON(address & ~HPAGE_PMD_MASK); 1129 1130 pmd = mm_find_pmd(mm, address); 1131 if (!pmd) { 1132 result = SCAN_PMD_NULL; 1133 goto out; 1134 } 1135 1136 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load)); 1137 pte = pte_offset_map_lock(mm, pmd, address, &ptl); 1138 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR; 1139 _pte++, _address += PAGE_SIZE) { 1140 pte_t pteval = *_pte; 1141 if (is_swap_pte(pteval)) { 1142 if (++unmapped <= khugepaged_max_ptes_swap) { 1143 continue; 1144 } else { 1145 result = SCAN_EXCEED_SWAP_PTE; 1146 goto out_unmap; 1147 } 1148 } 1149 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { 1150 if (!userfaultfd_armed(vma) && 1151 ++none_or_zero <= khugepaged_max_ptes_none) { 1152 continue; 1153 } else { 1154 result = SCAN_EXCEED_NONE_PTE; 1155 goto out_unmap; 1156 } 1157 } 1158 if (!pte_present(pteval)) { 1159 result = SCAN_PTE_NON_PRESENT; 1160 goto out_unmap; 1161 } 1162 if (pte_write(pteval)) 1163 writable = true; 1164 1165 page = vm_normal_page(vma, _address, pteval); 1166 if (unlikely(!page)) { 1167 result = SCAN_PAGE_NULL; 1168 goto out_unmap; 1169 } 1170 1171 /* TODO: teach khugepaged to collapse THP mapped with pte */ 1172 if (PageCompound(page)) { 1173 result = SCAN_PAGE_COMPOUND; 1174 goto out_unmap; 1175 } 1176 1177 /* 1178 * Record which node the original page is from and save this 1179 * information to khugepaged_node_load[]. 1180 * Khupaged will allocate hugepage from the node has the max 1181 * hit record. 1182 */ 1183 node = page_to_nid(page); 1184 if (khugepaged_scan_abort(node)) { 1185 result = SCAN_SCAN_ABORT; 1186 goto out_unmap; 1187 } 1188 khugepaged_node_load[node]++; 1189 if (!PageLRU(page)) { 1190 result = SCAN_PAGE_LRU; 1191 goto out_unmap; 1192 } 1193 if (PageLocked(page)) { 1194 result = SCAN_PAGE_LOCK; 1195 goto out_unmap; 1196 } 1197 if (!PageAnon(page)) { 1198 result = SCAN_PAGE_ANON; 1199 goto out_unmap; 1200 } 1201 1202 /* 1203 * cannot use mapcount: can't collapse if there's a gup pin. 1204 * The page must only be referenced by the scanned process 1205 * and page swap cache. 1206 */ 1207 if (page_count(page) != 1 + PageSwapCache(page)) { 1208 result = SCAN_PAGE_COUNT; 1209 goto out_unmap; 1210 } 1211 if (pte_young(pteval) || 1212 page_is_young(page) || PageReferenced(page) || 1213 mmu_notifier_test_young(vma->vm_mm, address)) 1214 referenced++; 1215 } 1216 if (writable) { 1217 if (referenced) { 1218 result = SCAN_SUCCEED; 1219 ret = 1; 1220 } else { 1221 result = SCAN_LACK_REFERENCED_PAGE; 1222 } 1223 } else { 1224 result = SCAN_PAGE_RO; 1225 } 1226 out_unmap: 1227 pte_unmap_unlock(pte, ptl); 1228 if (ret) { 1229 node = khugepaged_find_target_node(); 1230 /* collapse_huge_page will return with the mmap_sem released */ 1231 collapse_huge_page(mm, address, hpage, node, referenced); 1232 } 1233 out: 1234 trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced, 1235 none_or_zero, result, unmapped); 1236 return ret; 1237 } 1238 1239 static void collect_mm_slot(struct mm_slot *mm_slot) 1240 { 1241 struct mm_struct *mm = mm_slot->mm; 1242 1243 lockdep_assert_held(&khugepaged_mm_lock); 1244 1245 if (khugepaged_test_exit(mm)) { 1246 /* free mm_slot */ 1247 hash_del(&mm_slot->hash); 1248 list_del(&mm_slot->mm_node); 1249 1250 /* 1251 * Not strictly needed because the mm exited already. 1252 * 1253 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags); 1254 */ 1255 1256 /* khugepaged_mm_lock actually not necessary for the below */ 1257 free_mm_slot(mm_slot); 1258 mmdrop(mm); 1259 } 1260 } 1261 1262 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE) 1263 /* 1264 * Notify khugepaged that given addr of the mm is pte-mapped THP. Then 1265 * khugepaged should try to collapse the page table. 1266 */ 1267 static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm, 1268 unsigned long addr) 1269 { 1270 struct mm_slot *mm_slot; 1271 1272 VM_BUG_ON(addr & ~HPAGE_PMD_MASK); 1273 1274 spin_lock(&khugepaged_mm_lock); 1275 mm_slot = get_mm_slot(mm); 1276 if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP)) 1277 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr; 1278 spin_unlock(&khugepaged_mm_lock); 1279 return 0; 1280 } 1281 1282 /** 1283 * Try to collapse a pte-mapped THP for mm at address haddr. 1284 * 1285 * This function checks whether all the PTEs in the PMD are pointing to the 1286 * right THP. If so, retract the page table so the THP can refault in with 1287 * as pmd-mapped. 1288 */ 1289 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr) 1290 { 1291 unsigned long haddr = addr & HPAGE_PMD_MASK; 1292 struct vm_area_struct *vma = find_vma(mm, haddr); 1293 struct page *hpage = NULL; 1294 pte_t *start_pte, *pte; 1295 pmd_t *pmd, _pmd; 1296 spinlock_t *ptl; 1297 int count = 0; 1298 int i; 1299 1300 if (!vma || !vma->vm_file || 1301 vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE) 1302 return; 1303 1304 /* 1305 * This vm_flags may not have VM_HUGEPAGE if the page was not 1306 * collapsed by this mm. But we can still collapse if the page is 1307 * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check() 1308 * will not fail the vma for missing VM_HUGEPAGE 1309 */ 1310 if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE)) 1311 return; 1312 1313 pmd = mm_find_pmd(mm, haddr); 1314 if (!pmd) 1315 return; 1316 1317 start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl); 1318 1319 /* step 1: check all mapped PTEs are to the right huge page */ 1320 for (i = 0, addr = haddr, pte = start_pte; 1321 i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) { 1322 struct page *page; 1323 1324 /* empty pte, skip */ 1325 if (pte_none(*pte)) 1326 continue; 1327 1328 /* page swapped out, abort */ 1329 if (!pte_present(*pte)) 1330 goto abort; 1331 1332 page = vm_normal_page(vma, addr, *pte); 1333 1334 if (!page || !PageCompound(page)) 1335 goto abort; 1336 1337 if (!hpage) { 1338 hpage = compound_head(page); 1339 /* 1340 * The mapping of the THP should not change. 1341 * 1342 * Note that uprobe, debugger, or MAP_PRIVATE may 1343 * change the page table, but the new page will 1344 * not pass PageCompound() check. 1345 */ 1346 if (WARN_ON(hpage->mapping != vma->vm_file->f_mapping)) 1347 goto abort; 1348 } 1349 1350 /* 1351 * Confirm the page maps to the correct subpage. 1352 * 1353 * Note that uprobe, debugger, or MAP_PRIVATE may change 1354 * the page table, but the new page will not pass 1355 * PageCompound() check. 1356 */ 1357 if (WARN_ON(hpage + i != page)) 1358 goto abort; 1359 count++; 1360 } 1361 1362 /* step 2: adjust rmap */ 1363 for (i = 0, addr = haddr, pte = start_pte; 1364 i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) { 1365 struct page *page; 1366 1367 if (pte_none(*pte)) 1368 continue; 1369 page = vm_normal_page(vma, addr, *pte); 1370 page_remove_rmap(page, false); 1371 } 1372 1373 pte_unmap_unlock(start_pte, ptl); 1374 1375 /* step 3: set proper refcount and mm_counters. */ 1376 if (hpage) { 1377 page_ref_sub(hpage, count); 1378 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count); 1379 } 1380 1381 /* step 4: collapse pmd */ 1382 ptl = pmd_lock(vma->vm_mm, pmd); 1383 _pmd = pmdp_collapse_flush(vma, addr, pmd); 1384 spin_unlock(ptl); 1385 mm_dec_nr_ptes(mm); 1386 pte_free(mm, pmd_pgtable(_pmd)); 1387 return; 1388 1389 abort: 1390 pte_unmap_unlock(start_pte, ptl); 1391 } 1392 1393 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot) 1394 { 1395 struct mm_struct *mm = mm_slot->mm; 1396 int i; 1397 1398 if (likely(mm_slot->nr_pte_mapped_thp == 0)) 1399 return 0; 1400 1401 if (!down_write_trylock(&mm->mmap_sem)) 1402 return -EBUSY; 1403 1404 if (unlikely(khugepaged_test_exit(mm))) 1405 goto out; 1406 1407 for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++) 1408 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]); 1409 1410 out: 1411 mm_slot->nr_pte_mapped_thp = 0; 1412 up_write(&mm->mmap_sem); 1413 return 0; 1414 } 1415 1416 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff) 1417 { 1418 struct vm_area_struct *vma; 1419 unsigned long addr; 1420 pmd_t *pmd, _pmd; 1421 1422 i_mmap_lock_write(mapping); 1423 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) { 1424 /* 1425 * Check vma->anon_vma to exclude MAP_PRIVATE mappings that 1426 * got written to. These VMAs are likely not worth investing 1427 * down_write(mmap_sem) as PMD-mapping is likely to be split 1428 * later. 1429 * 1430 * Not that vma->anon_vma check is racy: it can be set up after 1431 * the check but before we took mmap_sem by the fault path. 1432 * But page lock would prevent establishing any new ptes of the 1433 * page, so we are safe. 1434 * 1435 * An alternative would be drop the check, but check that page 1436 * table is clear before calling pmdp_collapse_flush() under 1437 * ptl. It has higher chance to recover THP for the VMA, but 1438 * has higher cost too. 1439 */ 1440 if (vma->anon_vma) 1441 continue; 1442 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); 1443 if (addr & ~HPAGE_PMD_MASK) 1444 continue; 1445 if (vma->vm_end < addr + HPAGE_PMD_SIZE) 1446 continue; 1447 pmd = mm_find_pmd(vma->vm_mm, addr); 1448 if (!pmd) 1449 continue; 1450 /* 1451 * We need exclusive mmap_sem to retract page table. 1452 * 1453 * We use trylock due to lock inversion: we need to acquire 1454 * mmap_sem while holding page lock. Fault path does it in 1455 * reverse order. Trylock is a way to avoid deadlock. 1456 */ 1457 if (down_write_trylock(&vma->vm_mm->mmap_sem)) { 1458 spinlock_t *ptl = pmd_lock(vma->vm_mm, pmd); 1459 /* assume page table is clear */ 1460 _pmd = pmdp_collapse_flush(vma, addr, pmd); 1461 spin_unlock(ptl); 1462 up_write(&vma->vm_mm->mmap_sem); 1463 mm_dec_nr_ptes(vma->vm_mm); 1464 pte_free(vma->vm_mm, pmd_pgtable(_pmd)); 1465 } else { 1466 /* Try again later */ 1467 khugepaged_add_pte_mapped_thp(vma->vm_mm, addr); 1468 } 1469 } 1470 i_mmap_unlock_write(mapping); 1471 } 1472 1473 /** 1474 * collapse_file - collapse filemap/tmpfs/shmem pages into huge one. 1475 * 1476 * Basic scheme is simple, details are more complex: 1477 * - allocate and lock a new huge page; 1478 * - scan page cache replacing old pages with the new one 1479 * + swap/gup in pages if necessary; 1480 * + fill in gaps; 1481 * + keep old pages around in case rollback is required; 1482 * - if replacing succeeds: 1483 * + copy data over; 1484 * + free old pages; 1485 * + unlock huge page; 1486 * - if replacing failed; 1487 * + put all pages back and unfreeze them; 1488 * + restore gaps in the page cache; 1489 * + unlock and free huge page; 1490 */ 1491 static void collapse_file(struct mm_struct *mm, 1492 struct file *file, pgoff_t start, 1493 struct page **hpage, int node) 1494 { 1495 struct address_space *mapping = file->f_mapping; 1496 gfp_t gfp; 1497 struct page *new_page; 1498 struct mem_cgroup *memcg; 1499 pgoff_t index, end = start + HPAGE_PMD_NR; 1500 LIST_HEAD(pagelist); 1501 XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER); 1502 int nr_none = 0, result = SCAN_SUCCEED; 1503 bool is_shmem = shmem_file(file); 1504 1505 VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem); 1506 VM_BUG_ON(start & (HPAGE_PMD_NR - 1)); 1507 1508 /* Only allocate from the target node */ 1509 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE; 1510 1511 new_page = khugepaged_alloc_page(hpage, gfp, node); 1512 if (!new_page) { 1513 result = SCAN_ALLOC_HUGE_PAGE_FAIL; 1514 goto out; 1515 } 1516 1517 if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) { 1518 result = SCAN_CGROUP_CHARGE_FAIL; 1519 goto out; 1520 } 1521 1522 /* This will be less messy when we use multi-index entries */ 1523 do { 1524 xas_lock_irq(&xas); 1525 xas_create_range(&xas); 1526 if (!xas_error(&xas)) 1527 break; 1528 xas_unlock_irq(&xas); 1529 if (!xas_nomem(&xas, GFP_KERNEL)) { 1530 mem_cgroup_cancel_charge(new_page, memcg, true); 1531 result = SCAN_FAIL; 1532 goto out; 1533 } 1534 } while (1); 1535 1536 __SetPageLocked(new_page); 1537 if (is_shmem) 1538 __SetPageSwapBacked(new_page); 1539 new_page->index = start; 1540 new_page->mapping = mapping; 1541 1542 /* 1543 * At this point the new_page is locked and not up-to-date. 1544 * It's safe to insert it into the page cache, because nobody would 1545 * be able to map it or use it in another way until we unlock it. 1546 */ 1547 1548 xas_set(&xas, start); 1549 for (index = start; index < end; index++) { 1550 struct page *page = xas_next(&xas); 1551 1552 VM_BUG_ON(index != xas.xa_index); 1553 if (is_shmem) { 1554 if (!page) { 1555 /* 1556 * Stop if extent has been truncated or 1557 * hole-punched, and is now completely 1558 * empty. 1559 */ 1560 if (index == start) { 1561 if (!xas_next_entry(&xas, end - 1)) { 1562 result = SCAN_TRUNCATED; 1563 goto xa_locked; 1564 } 1565 xas_set(&xas, index); 1566 } 1567 if (!shmem_charge(mapping->host, 1)) { 1568 result = SCAN_FAIL; 1569 goto xa_locked; 1570 } 1571 xas_store(&xas, new_page); 1572 nr_none++; 1573 continue; 1574 } 1575 1576 if (xa_is_value(page) || !PageUptodate(page)) { 1577 xas_unlock_irq(&xas); 1578 /* swap in or instantiate fallocated page */ 1579 if (shmem_getpage(mapping->host, index, &page, 1580 SGP_NOHUGE)) { 1581 result = SCAN_FAIL; 1582 goto xa_unlocked; 1583 } 1584 } else if (trylock_page(page)) { 1585 get_page(page); 1586 xas_unlock_irq(&xas); 1587 } else { 1588 result = SCAN_PAGE_LOCK; 1589 goto xa_locked; 1590 } 1591 } else { /* !is_shmem */ 1592 if (!page || xa_is_value(page)) { 1593 xas_unlock_irq(&xas); 1594 page_cache_sync_readahead(mapping, &file->f_ra, 1595 file, index, 1596 PAGE_SIZE); 1597 /* drain pagevecs to help isolate_lru_page() */ 1598 lru_add_drain(); 1599 page = find_lock_page(mapping, index); 1600 if (unlikely(page == NULL)) { 1601 result = SCAN_FAIL; 1602 goto xa_unlocked; 1603 } 1604 } else if (!PageUptodate(page)) { 1605 xas_unlock_irq(&xas); 1606 wait_on_page_locked(page); 1607 if (!trylock_page(page)) { 1608 result = SCAN_PAGE_LOCK; 1609 goto xa_unlocked; 1610 } 1611 get_page(page); 1612 } else if (PageDirty(page)) { 1613 result = SCAN_FAIL; 1614 goto xa_locked; 1615 } else if (trylock_page(page)) { 1616 get_page(page); 1617 xas_unlock_irq(&xas); 1618 } else { 1619 result = SCAN_PAGE_LOCK; 1620 goto xa_locked; 1621 } 1622 } 1623 1624 /* 1625 * The page must be locked, so we can drop the i_pages lock 1626 * without racing with truncate. 1627 */ 1628 VM_BUG_ON_PAGE(!PageLocked(page), page); 1629 VM_BUG_ON_PAGE(!PageUptodate(page), page); 1630 1631 /* 1632 * If file was truncated then extended, or hole-punched, before 1633 * we locked the first page, then a THP might be there already. 1634 */ 1635 if (PageTransCompound(page)) { 1636 result = SCAN_PAGE_COMPOUND; 1637 goto out_unlock; 1638 } 1639 1640 if (page_mapping(page) != mapping) { 1641 result = SCAN_TRUNCATED; 1642 goto out_unlock; 1643 } 1644 1645 if (isolate_lru_page(page)) { 1646 result = SCAN_DEL_PAGE_LRU; 1647 goto out_unlock; 1648 } 1649 1650 if (page_has_private(page) && 1651 !try_to_release_page(page, GFP_KERNEL)) { 1652 result = SCAN_PAGE_HAS_PRIVATE; 1653 goto out_unlock; 1654 } 1655 1656 if (page_mapped(page)) 1657 unmap_mapping_pages(mapping, index, 1, false); 1658 1659 xas_lock_irq(&xas); 1660 xas_set(&xas, index); 1661 1662 VM_BUG_ON_PAGE(page != xas_load(&xas), page); 1663 VM_BUG_ON_PAGE(page_mapped(page), page); 1664 1665 /* 1666 * The page is expected to have page_count() == 3: 1667 * - we hold a pin on it; 1668 * - one reference from page cache; 1669 * - one from isolate_lru_page; 1670 */ 1671 if (!page_ref_freeze(page, 3)) { 1672 result = SCAN_PAGE_COUNT; 1673 xas_unlock_irq(&xas); 1674 putback_lru_page(page); 1675 goto out_unlock; 1676 } 1677 1678 /* 1679 * Add the page to the list to be able to undo the collapse if 1680 * something go wrong. 1681 */ 1682 list_add_tail(&page->lru, &pagelist); 1683 1684 /* Finally, replace with the new page. */ 1685 xas_store(&xas, new_page); 1686 continue; 1687 out_unlock: 1688 unlock_page(page); 1689 put_page(page); 1690 goto xa_unlocked; 1691 } 1692 1693 if (is_shmem) 1694 __inc_node_page_state(new_page, NR_SHMEM_THPS); 1695 else { 1696 __inc_node_page_state(new_page, NR_FILE_THPS); 1697 filemap_nr_thps_inc(mapping); 1698 } 1699 1700 if (nr_none) { 1701 struct zone *zone = page_zone(new_page); 1702 1703 __mod_node_page_state(zone->zone_pgdat, NR_FILE_PAGES, nr_none); 1704 if (is_shmem) 1705 __mod_node_page_state(zone->zone_pgdat, 1706 NR_SHMEM, nr_none); 1707 } 1708 1709 xa_locked: 1710 xas_unlock_irq(&xas); 1711 xa_unlocked: 1712 1713 if (result == SCAN_SUCCEED) { 1714 struct page *page, *tmp; 1715 1716 /* 1717 * Replacing old pages with new one has succeeded, now we 1718 * need to copy the content and free the old pages. 1719 */ 1720 index = start; 1721 list_for_each_entry_safe(page, tmp, &pagelist, lru) { 1722 while (index < page->index) { 1723 clear_highpage(new_page + (index % HPAGE_PMD_NR)); 1724 index++; 1725 } 1726 copy_highpage(new_page + (page->index % HPAGE_PMD_NR), 1727 page); 1728 list_del(&page->lru); 1729 page->mapping = NULL; 1730 page_ref_unfreeze(page, 1); 1731 ClearPageActive(page); 1732 ClearPageUnevictable(page); 1733 unlock_page(page); 1734 put_page(page); 1735 index++; 1736 } 1737 while (index < end) { 1738 clear_highpage(new_page + (index % HPAGE_PMD_NR)); 1739 index++; 1740 } 1741 1742 SetPageUptodate(new_page); 1743 page_ref_add(new_page, HPAGE_PMD_NR - 1); 1744 mem_cgroup_commit_charge(new_page, memcg, false, true); 1745 1746 if (is_shmem) { 1747 set_page_dirty(new_page); 1748 lru_cache_add_anon(new_page); 1749 } else { 1750 lru_cache_add_file(new_page); 1751 } 1752 count_memcg_events(memcg, THP_COLLAPSE_ALLOC, 1); 1753 1754 /* 1755 * Remove pte page tables, so we can re-fault the page as huge. 1756 */ 1757 retract_page_tables(mapping, start); 1758 *hpage = NULL; 1759 1760 khugepaged_pages_collapsed++; 1761 } else { 1762 struct page *page; 1763 1764 /* Something went wrong: roll back page cache changes */ 1765 xas_lock_irq(&xas); 1766 mapping->nrpages -= nr_none; 1767 1768 if (is_shmem) 1769 shmem_uncharge(mapping->host, nr_none); 1770 1771 xas_set(&xas, start); 1772 xas_for_each(&xas, page, end - 1) { 1773 page = list_first_entry_or_null(&pagelist, 1774 struct page, lru); 1775 if (!page || xas.xa_index < page->index) { 1776 if (!nr_none) 1777 break; 1778 nr_none--; 1779 /* Put holes back where they were */ 1780 xas_store(&xas, NULL); 1781 continue; 1782 } 1783 1784 VM_BUG_ON_PAGE(page->index != xas.xa_index, page); 1785 1786 /* Unfreeze the page. */ 1787 list_del(&page->lru); 1788 page_ref_unfreeze(page, 2); 1789 xas_store(&xas, page); 1790 xas_pause(&xas); 1791 xas_unlock_irq(&xas); 1792 unlock_page(page); 1793 putback_lru_page(page); 1794 xas_lock_irq(&xas); 1795 } 1796 VM_BUG_ON(nr_none); 1797 xas_unlock_irq(&xas); 1798 1799 mem_cgroup_cancel_charge(new_page, memcg, true); 1800 new_page->mapping = NULL; 1801 } 1802 1803 unlock_page(new_page); 1804 out: 1805 VM_BUG_ON(!list_empty(&pagelist)); 1806 /* TODO: tracepoints */ 1807 } 1808 1809 static void khugepaged_scan_file(struct mm_struct *mm, 1810 struct file *file, pgoff_t start, struct page **hpage) 1811 { 1812 struct page *page = NULL; 1813 struct address_space *mapping = file->f_mapping; 1814 XA_STATE(xas, &mapping->i_pages, start); 1815 int present, swap; 1816 int node = NUMA_NO_NODE; 1817 int result = SCAN_SUCCEED; 1818 1819 present = 0; 1820 swap = 0; 1821 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load)); 1822 rcu_read_lock(); 1823 xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) { 1824 if (xas_retry(&xas, page)) 1825 continue; 1826 1827 if (xa_is_value(page)) { 1828 if (++swap > khugepaged_max_ptes_swap) { 1829 result = SCAN_EXCEED_SWAP_PTE; 1830 break; 1831 } 1832 continue; 1833 } 1834 1835 if (PageTransCompound(page)) { 1836 result = SCAN_PAGE_COMPOUND; 1837 break; 1838 } 1839 1840 node = page_to_nid(page); 1841 if (khugepaged_scan_abort(node)) { 1842 result = SCAN_SCAN_ABORT; 1843 break; 1844 } 1845 khugepaged_node_load[node]++; 1846 1847 if (!PageLRU(page)) { 1848 result = SCAN_PAGE_LRU; 1849 break; 1850 } 1851 1852 if (page_count(page) != 1853 1 + page_mapcount(page) + page_has_private(page)) { 1854 result = SCAN_PAGE_COUNT; 1855 break; 1856 } 1857 1858 /* 1859 * We probably should check if the page is referenced here, but 1860 * nobody would transfer pte_young() to PageReferenced() for us. 1861 * And rmap walk here is just too costly... 1862 */ 1863 1864 present++; 1865 1866 if (need_resched()) { 1867 xas_pause(&xas); 1868 cond_resched_rcu(); 1869 } 1870 } 1871 rcu_read_unlock(); 1872 1873 if (result == SCAN_SUCCEED) { 1874 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) { 1875 result = SCAN_EXCEED_NONE_PTE; 1876 } else { 1877 node = khugepaged_find_target_node(); 1878 collapse_file(mm, file, start, hpage, node); 1879 } 1880 } 1881 1882 /* TODO: tracepoints */ 1883 } 1884 #else 1885 static void khugepaged_scan_file(struct mm_struct *mm, 1886 struct file *file, pgoff_t start, struct page **hpage) 1887 { 1888 BUILD_BUG(); 1889 } 1890 1891 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot) 1892 { 1893 return 0; 1894 } 1895 #endif 1896 1897 static unsigned int khugepaged_scan_mm_slot(unsigned int pages, 1898 struct page **hpage) 1899 __releases(&khugepaged_mm_lock) 1900 __acquires(&khugepaged_mm_lock) 1901 { 1902 struct mm_slot *mm_slot; 1903 struct mm_struct *mm; 1904 struct vm_area_struct *vma; 1905 int progress = 0; 1906 1907 VM_BUG_ON(!pages); 1908 lockdep_assert_held(&khugepaged_mm_lock); 1909 1910 if (khugepaged_scan.mm_slot) 1911 mm_slot = khugepaged_scan.mm_slot; 1912 else { 1913 mm_slot = list_entry(khugepaged_scan.mm_head.next, 1914 struct mm_slot, mm_node); 1915 khugepaged_scan.address = 0; 1916 khugepaged_scan.mm_slot = mm_slot; 1917 } 1918 spin_unlock(&khugepaged_mm_lock); 1919 khugepaged_collapse_pte_mapped_thps(mm_slot); 1920 1921 mm = mm_slot->mm; 1922 /* 1923 * Don't wait for semaphore (to avoid long wait times). Just move to 1924 * the next mm on the list. 1925 */ 1926 vma = NULL; 1927 if (unlikely(!down_read_trylock(&mm->mmap_sem))) 1928 goto breakouterloop_mmap_sem; 1929 if (likely(!khugepaged_test_exit(mm))) 1930 vma = find_vma(mm, khugepaged_scan.address); 1931 1932 progress++; 1933 for (; vma; vma = vma->vm_next) { 1934 unsigned long hstart, hend; 1935 1936 cond_resched(); 1937 if (unlikely(khugepaged_test_exit(mm))) { 1938 progress++; 1939 break; 1940 } 1941 if (!hugepage_vma_check(vma, vma->vm_flags)) { 1942 skip: 1943 progress++; 1944 continue; 1945 } 1946 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; 1947 hend = vma->vm_end & HPAGE_PMD_MASK; 1948 if (hstart >= hend) 1949 goto skip; 1950 if (khugepaged_scan.address > hend) 1951 goto skip; 1952 if (khugepaged_scan.address < hstart) 1953 khugepaged_scan.address = hstart; 1954 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK); 1955 1956 while (khugepaged_scan.address < hend) { 1957 int ret; 1958 cond_resched(); 1959 if (unlikely(khugepaged_test_exit(mm))) 1960 goto breakouterloop; 1961 1962 VM_BUG_ON(khugepaged_scan.address < hstart || 1963 khugepaged_scan.address + HPAGE_PMD_SIZE > 1964 hend); 1965 if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) { 1966 struct file *file; 1967 pgoff_t pgoff = linear_page_index(vma, 1968 khugepaged_scan.address); 1969 1970 if (shmem_file(vma->vm_file) 1971 && !shmem_huge_enabled(vma)) 1972 goto skip; 1973 file = get_file(vma->vm_file); 1974 up_read(&mm->mmap_sem); 1975 ret = 1; 1976 khugepaged_scan_file(mm, file, pgoff, hpage); 1977 fput(file); 1978 } else { 1979 ret = khugepaged_scan_pmd(mm, vma, 1980 khugepaged_scan.address, 1981 hpage); 1982 } 1983 /* move to next address */ 1984 khugepaged_scan.address += HPAGE_PMD_SIZE; 1985 progress += HPAGE_PMD_NR; 1986 if (ret) 1987 /* we released mmap_sem so break loop */ 1988 goto breakouterloop_mmap_sem; 1989 if (progress >= pages) 1990 goto breakouterloop; 1991 } 1992 } 1993 breakouterloop: 1994 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */ 1995 breakouterloop_mmap_sem: 1996 1997 spin_lock(&khugepaged_mm_lock); 1998 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot); 1999 /* 2000 * Release the current mm_slot if this mm is about to die, or 2001 * if we scanned all vmas of this mm. 2002 */ 2003 if (khugepaged_test_exit(mm) || !vma) { 2004 /* 2005 * Make sure that if mm_users is reaching zero while 2006 * khugepaged runs here, khugepaged_exit will find 2007 * mm_slot not pointing to the exiting mm. 2008 */ 2009 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) { 2010 khugepaged_scan.mm_slot = list_entry( 2011 mm_slot->mm_node.next, 2012 struct mm_slot, mm_node); 2013 khugepaged_scan.address = 0; 2014 } else { 2015 khugepaged_scan.mm_slot = NULL; 2016 khugepaged_full_scans++; 2017 } 2018 2019 collect_mm_slot(mm_slot); 2020 } 2021 2022 return progress; 2023 } 2024 2025 static int khugepaged_has_work(void) 2026 { 2027 return !list_empty(&khugepaged_scan.mm_head) && 2028 khugepaged_enabled(); 2029 } 2030 2031 static int khugepaged_wait_event(void) 2032 { 2033 return !list_empty(&khugepaged_scan.mm_head) || 2034 kthread_should_stop(); 2035 } 2036 2037 static void khugepaged_do_scan(void) 2038 { 2039 struct page *hpage = NULL; 2040 unsigned int progress = 0, pass_through_head = 0; 2041 unsigned int pages = khugepaged_pages_to_scan; 2042 bool wait = true; 2043 2044 barrier(); /* write khugepaged_pages_to_scan to local stack */ 2045 2046 while (progress < pages) { 2047 if (!khugepaged_prealloc_page(&hpage, &wait)) 2048 break; 2049 2050 cond_resched(); 2051 2052 if (unlikely(kthread_should_stop() || try_to_freeze())) 2053 break; 2054 2055 spin_lock(&khugepaged_mm_lock); 2056 if (!khugepaged_scan.mm_slot) 2057 pass_through_head++; 2058 if (khugepaged_has_work() && 2059 pass_through_head < 2) 2060 progress += khugepaged_scan_mm_slot(pages - progress, 2061 &hpage); 2062 else 2063 progress = pages; 2064 spin_unlock(&khugepaged_mm_lock); 2065 } 2066 2067 if (!IS_ERR_OR_NULL(hpage)) 2068 put_page(hpage); 2069 } 2070 2071 static bool khugepaged_should_wakeup(void) 2072 { 2073 return kthread_should_stop() || 2074 time_after_eq(jiffies, khugepaged_sleep_expire); 2075 } 2076 2077 static void khugepaged_wait_work(void) 2078 { 2079 if (khugepaged_has_work()) { 2080 const unsigned long scan_sleep_jiffies = 2081 msecs_to_jiffies(khugepaged_scan_sleep_millisecs); 2082 2083 if (!scan_sleep_jiffies) 2084 return; 2085 2086 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies; 2087 wait_event_freezable_timeout(khugepaged_wait, 2088 khugepaged_should_wakeup(), 2089 scan_sleep_jiffies); 2090 return; 2091 } 2092 2093 if (khugepaged_enabled()) 2094 wait_event_freezable(khugepaged_wait, khugepaged_wait_event()); 2095 } 2096 2097 static int khugepaged(void *none) 2098 { 2099 struct mm_slot *mm_slot; 2100 2101 set_freezable(); 2102 set_user_nice(current, MAX_NICE); 2103 2104 while (!kthread_should_stop()) { 2105 khugepaged_do_scan(); 2106 khugepaged_wait_work(); 2107 } 2108 2109 spin_lock(&khugepaged_mm_lock); 2110 mm_slot = khugepaged_scan.mm_slot; 2111 khugepaged_scan.mm_slot = NULL; 2112 if (mm_slot) 2113 collect_mm_slot(mm_slot); 2114 spin_unlock(&khugepaged_mm_lock); 2115 return 0; 2116 } 2117 2118 static void set_recommended_min_free_kbytes(void) 2119 { 2120 struct zone *zone; 2121 int nr_zones = 0; 2122 unsigned long recommended_min; 2123 2124 for_each_populated_zone(zone) { 2125 /* 2126 * We don't need to worry about fragmentation of 2127 * ZONE_MOVABLE since it only has movable pages. 2128 */ 2129 if (zone_idx(zone) > gfp_zone(GFP_USER)) 2130 continue; 2131 2132 nr_zones++; 2133 } 2134 2135 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */ 2136 recommended_min = pageblock_nr_pages * nr_zones * 2; 2137 2138 /* 2139 * Make sure that on average at least two pageblocks are almost free 2140 * of another type, one for a migratetype to fall back to and a 2141 * second to avoid subsequent fallbacks of other types There are 3 2142 * MIGRATE_TYPES we care about. 2143 */ 2144 recommended_min += pageblock_nr_pages * nr_zones * 2145 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES; 2146 2147 /* don't ever allow to reserve more than 5% of the lowmem */ 2148 recommended_min = min(recommended_min, 2149 (unsigned long) nr_free_buffer_pages() / 20); 2150 recommended_min <<= (PAGE_SHIFT-10); 2151 2152 if (recommended_min > min_free_kbytes) { 2153 if (user_min_free_kbytes >= 0) 2154 pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n", 2155 min_free_kbytes, recommended_min); 2156 2157 min_free_kbytes = recommended_min; 2158 } 2159 setup_per_zone_wmarks(); 2160 } 2161 2162 int start_stop_khugepaged(void) 2163 { 2164 static struct task_struct *khugepaged_thread __read_mostly; 2165 static DEFINE_MUTEX(khugepaged_mutex); 2166 int err = 0; 2167 2168 mutex_lock(&khugepaged_mutex); 2169 if (khugepaged_enabled()) { 2170 if (!khugepaged_thread) 2171 khugepaged_thread = kthread_run(khugepaged, NULL, 2172 "khugepaged"); 2173 if (IS_ERR(khugepaged_thread)) { 2174 pr_err("khugepaged: kthread_run(khugepaged) failed\n"); 2175 err = PTR_ERR(khugepaged_thread); 2176 khugepaged_thread = NULL; 2177 goto fail; 2178 } 2179 2180 if (!list_empty(&khugepaged_scan.mm_head)) 2181 wake_up_interruptible(&khugepaged_wait); 2182 2183 set_recommended_min_free_kbytes(); 2184 } else if (khugepaged_thread) { 2185 kthread_stop(khugepaged_thread); 2186 khugepaged_thread = NULL; 2187 } 2188 fail: 2189 mutex_unlock(&khugepaged_mutex); 2190 return err; 2191 } 2192