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