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