1 /* 2 * Copyright (C) 2009 Red Hat, Inc. 3 * 4 * This work is licensed under the terms of the GNU GPL, version 2. See 5 * the COPYING file in the top-level directory. 6 */ 7 8 #include <linux/mm.h> 9 #include <linux/sched.h> 10 #include <linux/highmem.h> 11 #include <linux/hugetlb.h> 12 #include <linux/mmu_notifier.h> 13 #include <linux/rmap.h> 14 #include <linux/swap.h> 15 #include <linux/mm_inline.h> 16 #include <linux/kthread.h> 17 #include <linux/khugepaged.h> 18 #include <linux/freezer.h> 19 #include <linux/mman.h> 20 #include <asm/tlb.h> 21 #include <asm/pgalloc.h> 22 #include "internal.h" 23 24 /* 25 * By default transparent hugepage support is enabled for all mappings 26 * and khugepaged scans all mappings. Defrag is only invoked by 27 * khugepaged hugepage allocations and by page faults inside 28 * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived 29 * allocations. 30 */ 31 unsigned long transparent_hugepage_flags __read_mostly = 32 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS 33 (1<<TRANSPARENT_HUGEPAGE_FLAG)| 34 #endif 35 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE 36 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)| 37 #endif 38 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)| 39 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); 40 41 /* default scan 8*512 pte (or vmas) every 30 second */ 42 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8; 43 static unsigned int khugepaged_pages_collapsed; 44 static unsigned int khugepaged_full_scans; 45 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000; 46 /* during fragmentation poll the hugepage allocator once every minute */ 47 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000; 48 static struct task_struct *khugepaged_thread __read_mostly; 49 static DEFINE_MUTEX(khugepaged_mutex); 50 static DEFINE_SPINLOCK(khugepaged_mm_lock); 51 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait); 52 /* 53 * default collapse hugepages if there is at least one pte mapped like 54 * it would have happened if the vma was large enough during page 55 * fault. 56 */ 57 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1; 58 59 static int khugepaged(void *none); 60 static int mm_slots_hash_init(void); 61 static int khugepaged_slab_init(void); 62 static void khugepaged_slab_free(void); 63 64 #define MM_SLOTS_HASH_HEADS 1024 65 static struct hlist_head *mm_slots_hash __read_mostly; 66 static struct kmem_cache *mm_slot_cache __read_mostly; 67 68 /** 69 * struct mm_slot - hash lookup from mm to mm_slot 70 * @hash: hash collision list 71 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head 72 * @mm: the mm that this information is valid for 73 */ 74 struct mm_slot { 75 struct hlist_node hash; 76 struct list_head mm_node; 77 struct mm_struct *mm; 78 }; 79 80 /** 81 * struct khugepaged_scan - cursor for scanning 82 * @mm_head: the head of the mm list to scan 83 * @mm_slot: the current mm_slot we are scanning 84 * @address: the next address inside that to be scanned 85 * 86 * There is only the one khugepaged_scan instance of this cursor structure. 87 */ 88 struct khugepaged_scan { 89 struct list_head mm_head; 90 struct mm_slot *mm_slot; 91 unsigned long address; 92 }; 93 static struct khugepaged_scan khugepaged_scan = { 94 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head), 95 }; 96 97 98 static int set_recommended_min_free_kbytes(void) 99 { 100 struct zone *zone; 101 int nr_zones = 0; 102 unsigned long recommended_min; 103 extern int min_free_kbytes; 104 105 if (!khugepaged_enabled()) 106 return 0; 107 108 for_each_populated_zone(zone) 109 nr_zones++; 110 111 /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */ 112 recommended_min = pageblock_nr_pages * nr_zones * 2; 113 114 /* 115 * Make sure that on average at least two pageblocks are almost free 116 * of another type, one for a migratetype to fall back to and a 117 * second to avoid subsequent fallbacks of other types There are 3 118 * MIGRATE_TYPES we care about. 119 */ 120 recommended_min += pageblock_nr_pages * nr_zones * 121 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES; 122 123 /* don't ever allow to reserve more than 5% of the lowmem */ 124 recommended_min = min(recommended_min, 125 (unsigned long) nr_free_buffer_pages() / 20); 126 recommended_min <<= (PAGE_SHIFT-10); 127 128 if (recommended_min > min_free_kbytes) 129 min_free_kbytes = recommended_min; 130 setup_per_zone_wmarks(); 131 return 0; 132 } 133 late_initcall(set_recommended_min_free_kbytes); 134 135 static int start_khugepaged(void) 136 { 137 int err = 0; 138 if (khugepaged_enabled()) { 139 if (!khugepaged_thread) 140 khugepaged_thread = kthread_run(khugepaged, NULL, 141 "khugepaged"); 142 if (unlikely(IS_ERR(khugepaged_thread))) { 143 printk(KERN_ERR 144 "khugepaged: kthread_run(khugepaged) failed\n"); 145 err = PTR_ERR(khugepaged_thread); 146 khugepaged_thread = NULL; 147 } 148 149 if (!list_empty(&khugepaged_scan.mm_head)) 150 wake_up_interruptible(&khugepaged_wait); 151 152 set_recommended_min_free_kbytes(); 153 } else if (khugepaged_thread) { 154 kthread_stop(khugepaged_thread); 155 khugepaged_thread = NULL; 156 } 157 158 return err; 159 } 160 161 #ifdef CONFIG_SYSFS 162 163 static ssize_t double_flag_show(struct kobject *kobj, 164 struct kobj_attribute *attr, char *buf, 165 enum transparent_hugepage_flag enabled, 166 enum transparent_hugepage_flag req_madv) 167 { 168 if (test_bit(enabled, &transparent_hugepage_flags)) { 169 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags)); 170 return sprintf(buf, "[always] madvise never\n"); 171 } else if (test_bit(req_madv, &transparent_hugepage_flags)) 172 return sprintf(buf, "always [madvise] never\n"); 173 else 174 return sprintf(buf, "always madvise [never]\n"); 175 } 176 static ssize_t double_flag_store(struct kobject *kobj, 177 struct kobj_attribute *attr, 178 const char *buf, size_t count, 179 enum transparent_hugepage_flag enabled, 180 enum transparent_hugepage_flag req_madv) 181 { 182 if (!memcmp("always", buf, 183 min(sizeof("always")-1, count))) { 184 set_bit(enabled, &transparent_hugepage_flags); 185 clear_bit(req_madv, &transparent_hugepage_flags); 186 } else if (!memcmp("madvise", buf, 187 min(sizeof("madvise")-1, count))) { 188 clear_bit(enabled, &transparent_hugepage_flags); 189 set_bit(req_madv, &transparent_hugepage_flags); 190 } else if (!memcmp("never", buf, 191 min(sizeof("never")-1, count))) { 192 clear_bit(enabled, &transparent_hugepage_flags); 193 clear_bit(req_madv, &transparent_hugepage_flags); 194 } else 195 return -EINVAL; 196 197 return count; 198 } 199 200 static ssize_t enabled_show(struct kobject *kobj, 201 struct kobj_attribute *attr, char *buf) 202 { 203 return double_flag_show(kobj, attr, buf, 204 TRANSPARENT_HUGEPAGE_FLAG, 205 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG); 206 } 207 static ssize_t enabled_store(struct kobject *kobj, 208 struct kobj_attribute *attr, 209 const char *buf, size_t count) 210 { 211 ssize_t ret; 212 213 ret = double_flag_store(kobj, attr, buf, count, 214 TRANSPARENT_HUGEPAGE_FLAG, 215 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG); 216 217 if (ret > 0) { 218 int err; 219 220 mutex_lock(&khugepaged_mutex); 221 err = start_khugepaged(); 222 mutex_unlock(&khugepaged_mutex); 223 224 if (err) 225 ret = err; 226 } 227 228 return ret; 229 } 230 static struct kobj_attribute enabled_attr = 231 __ATTR(enabled, 0644, enabled_show, enabled_store); 232 233 static ssize_t single_flag_show(struct kobject *kobj, 234 struct kobj_attribute *attr, char *buf, 235 enum transparent_hugepage_flag flag) 236 { 237 return sprintf(buf, "%d\n", 238 !!test_bit(flag, &transparent_hugepage_flags)); 239 } 240 241 static ssize_t single_flag_store(struct kobject *kobj, 242 struct kobj_attribute *attr, 243 const char *buf, size_t count, 244 enum transparent_hugepage_flag flag) 245 { 246 unsigned long value; 247 int ret; 248 249 ret = kstrtoul(buf, 10, &value); 250 if (ret < 0) 251 return ret; 252 if (value > 1) 253 return -EINVAL; 254 255 if (value) 256 set_bit(flag, &transparent_hugepage_flags); 257 else 258 clear_bit(flag, &transparent_hugepage_flags); 259 260 return count; 261 } 262 263 /* 264 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind 265 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of 266 * memory just to allocate one more hugepage. 267 */ 268 static ssize_t defrag_show(struct kobject *kobj, 269 struct kobj_attribute *attr, char *buf) 270 { 271 return double_flag_show(kobj, attr, buf, 272 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG, 273 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG); 274 } 275 static ssize_t defrag_store(struct kobject *kobj, 276 struct kobj_attribute *attr, 277 const char *buf, size_t count) 278 { 279 return double_flag_store(kobj, attr, buf, count, 280 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG, 281 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG); 282 } 283 static struct kobj_attribute defrag_attr = 284 __ATTR(defrag, 0644, defrag_show, defrag_store); 285 286 #ifdef CONFIG_DEBUG_VM 287 static ssize_t debug_cow_show(struct kobject *kobj, 288 struct kobj_attribute *attr, char *buf) 289 { 290 return single_flag_show(kobj, attr, buf, 291 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); 292 } 293 static ssize_t debug_cow_store(struct kobject *kobj, 294 struct kobj_attribute *attr, 295 const char *buf, size_t count) 296 { 297 return single_flag_store(kobj, attr, buf, count, 298 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); 299 } 300 static struct kobj_attribute debug_cow_attr = 301 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store); 302 #endif /* CONFIG_DEBUG_VM */ 303 304 static struct attribute *hugepage_attr[] = { 305 &enabled_attr.attr, 306 &defrag_attr.attr, 307 #ifdef CONFIG_DEBUG_VM 308 &debug_cow_attr.attr, 309 #endif 310 NULL, 311 }; 312 313 static struct attribute_group hugepage_attr_group = { 314 .attrs = hugepage_attr, 315 }; 316 317 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj, 318 struct kobj_attribute *attr, 319 char *buf) 320 { 321 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs); 322 } 323 324 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj, 325 struct kobj_attribute *attr, 326 const char *buf, size_t count) 327 { 328 unsigned long msecs; 329 int err; 330 331 err = strict_strtoul(buf, 10, &msecs); 332 if (err || msecs > UINT_MAX) 333 return -EINVAL; 334 335 khugepaged_scan_sleep_millisecs = msecs; 336 wake_up_interruptible(&khugepaged_wait); 337 338 return count; 339 } 340 static struct kobj_attribute scan_sleep_millisecs_attr = 341 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show, 342 scan_sleep_millisecs_store); 343 344 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj, 345 struct kobj_attribute *attr, 346 char *buf) 347 { 348 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs); 349 } 350 351 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj, 352 struct kobj_attribute *attr, 353 const char *buf, size_t count) 354 { 355 unsigned long msecs; 356 int err; 357 358 err = strict_strtoul(buf, 10, &msecs); 359 if (err || msecs > UINT_MAX) 360 return -EINVAL; 361 362 khugepaged_alloc_sleep_millisecs = msecs; 363 wake_up_interruptible(&khugepaged_wait); 364 365 return count; 366 } 367 static struct kobj_attribute alloc_sleep_millisecs_attr = 368 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show, 369 alloc_sleep_millisecs_store); 370 371 static ssize_t pages_to_scan_show(struct kobject *kobj, 372 struct kobj_attribute *attr, 373 char *buf) 374 { 375 return sprintf(buf, "%u\n", khugepaged_pages_to_scan); 376 } 377 static ssize_t pages_to_scan_store(struct kobject *kobj, 378 struct kobj_attribute *attr, 379 const char *buf, size_t count) 380 { 381 int err; 382 unsigned long pages; 383 384 err = strict_strtoul(buf, 10, &pages); 385 if (err || !pages || pages > UINT_MAX) 386 return -EINVAL; 387 388 khugepaged_pages_to_scan = pages; 389 390 return count; 391 } 392 static struct kobj_attribute pages_to_scan_attr = 393 __ATTR(pages_to_scan, 0644, pages_to_scan_show, 394 pages_to_scan_store); 395 396 static ssize_t pages_collapsed_show(struct kobject *kobj, 397 struct kobj_attribute *attr, 398 char *buf) 399 { 400 return sprintf(buf, "%u\n", khugepaged_pages_collapsed); 401 } 402 static struct kobj_attribute pages_collapsed_attr = 403 __ATTR_RO(pages_collapsed); 404 405 static ssize_t full_scans_show(struct kobject *kobj, 406 struct kobj_attribute *attr, 407 char *buf) 408 { 409 return sprintf(buf, "%u\n", khugepaged_full_scans); 410 } 411 static struct kobj_attribute full_scans_attr = 412 __ATTR_RO(full_scans); 413 414 static ssize_t khugepaged_defrag_show(struct kobject *kobj, 415 struct kobj_attribute *attr, char *buf) 416 { 417 return single_flag_show(kobj, attr, buf, 418 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); 419 } 420 static ssize_t khugepaged_defrag_store(struct kobject *kobj, 421 struct kobj_attribute *attr, 422 const char *buf, size_t count) 423 { 424 return single_flag_store(kobj, attr, buf, count, 425 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); 426 } 427 static struct kobj_attribute khugepaged_defrag_attr = 428 __ATTR(defrag, 0644, khugepaged_defrag_show, 429 khugepaged_defrag_store); 430 431 /* 432 * max_ptes_none controls if khugepaged should collapse hugepages over 433 * any unmapped ptes in turn potentially increasing the memory 434 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not 435 * reduce the available free memory in the system as it 436 * runs. Increasing max_ptes_none will instead potentially reduce the 437 * free memory in the system during the khugepaged scan. 438 */ 439 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj, 440 struct kobj_attribute *attr, 441 char *buf) 442 { 443 return sprintf(buf, "%u\n", khugepaged_max_ptes_none); 444 } 445 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj, 446 struct kobj_attribute *attr, 447 const char *buf, size_t count) 448 { 449 int err; 450 unsigned long max_ptes_none; 451 452 err = strict_strtoul(buf, 10, &max_ptes_none); 453 if (err || max_ptes_none > HPAGE_PMD_NR-1) 454 return -EINVAL; 455 456 khugepaged_max_ptes_none = max_ptes_none; 457 458 return count; 459 } 460 static struct kobj_attribute khugepaged_max_ptes_none_attr = 461 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show, 462 khugepaged_max_ptes_none_store); 463 464 static struct attribute *khugepaged_attr[] = { 465 &khugepaged_defrag_attr.attr, 466 &khugepaged_max_ptes_none_attr.attr, 467 &pages_to_scan_attr.attr, 468 &pages_collapsed_attr.attr, 469 &full_scans_attr.attr, 470 &scan_sleep_millisecs_attr.attr, 471 &alloc_sleep_millisecs_attr.attr, 472 NULL, 473 }; 474 475 static struct attribute_group khugepaged_attr_group = { 476 .attrs = khugepaged_attr, 477 .name = "khugepaged", 478 }; 479 480 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj) 481 { 482 int err; 483 484 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj); 485 if (unlikely(!*hugepage_kobj)) { 486 printk(KERN_ERR "hugepage: failed kobject create\n"); 487 return -ENOMEM; 488 } 489 490 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group); 491 if (err) { 492 printk(KERN_ERR "hugepage: failed register hugeage group\n"); 493 goto delete_obj; 494 } 495 496 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group); 497 if (err) { 498 printk(KERN_ERR "hugepage: failed register hugeage group\n"); 499 goto remove_hp_group; 500 } 501 502 return 0; 503 504 remove_hp_group: 505 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group); 506 delete_obj: 507 kobject_put(*hugepage_kobj); 508 return err; 509 } 510 511 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj) 512 { 513 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group); 514 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group); 515 kobject_put(hugepage_kobj); 516 } 517 #else 518 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj) 519 { 520 return 0; 521 } 522 523 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj) 524 { 525 } 526 #endif /* CONFIG_SYSFS */ 527 528 static int __init hugepage_init(void) 529 { 530 int err; 531 struct kobject *hugepage_kobj; 532 533 if (!has_transparent_hugepage()) { 534 transparent_hugepage_flags = 0; 535 return -EINVAL; 536 } 537 538 err = hugepage_init_sysfs(&hugepage_kobj); 539 if (err) 540 return err; 541 542 err = khugepaged_slab_init(); 543 if (err) 544 goto out; 545 546 err = mm_slots_hash_init(); 547 if (err) { 548 khugepaged_slab_free(); 549 goto out; 550 } 551 552 /* 553 * By default disable transparent hugepages on smaller systems, 554 * where the extra memory used could hurt more than TLB overhead 555 * is likely to save. The admin can still enable it through /sys. 556 */ 557 if (totalram_pages < (512 << (20 - PAGE_SHIFT))) 558 transparent_hugepage_flags = 0; 559 560 start_khugepaged(); 561 562 return 0; 563 out: 564 hugepage_exit_sysfs(hugepage_kobj); 565 return err; 566 } 567 module_init(hugepage_init) 568 569 static int __init setup_transparent_hugepage(char *str) 570 { 571 int ret = 0; 572 if (!str) 573 goto out; 574 if (!strcmp(str, "always")) { 575 set_bit(TRANSPARENT_HUGEPAGE_FLAG, 576 &transparent_hugepage_flags); 577 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 578 &transparent_hugepage_flags); 579 ret = 1; 580 } else if (!strcmp(str, "madvise")) { 581 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 582 &transparent_hugepage_flags); 583 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 584 &transparent_hugepage_flags); 585 ret = 1; 586 } else if (!strcmp(str, "never")) { 587 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 588 &transparent_hugepage_flags); 589 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 590 &transparent_hugepage_flags); 591 ret = 1; 592 } 593 out: 594 if (!ret) 595 printk(KERN_WARNING 596 "transparent_hugepage= cannot parse, ignored\n"); 597 return ret; 598 } 599 __setup("transparent_hugepage=", setup_transparent_hugepage); 600 601 static inline pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma) 602 { 603 if (likely(vma->vm_flags & VM_WRITE)) 604 pmd = pmd_mkwrite(pmd); 605 return pmd; 606 } 607 608 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm, 609 struct vm_area_struct *vma, 610 unsigned long haddr, pmd_t *pmd, 611 struct page *page) 612 { 613 pgtable_t pgtable; 614 615 VM_BUG_ON(!PageCompound(page)); 616 pgtable = pte_alloc_one(mm, haddr); 617 if (unlikely(!pgtable)) 618 return VM_FAULT_OOM; 619 620 clear_huge_page(page, haddr, HPAGE_PMD_NR); 621 __SetPageUptodate(page); 622 623 spin_lock(&mm->page_table_lock); 624 if (unlikely(!pmd_none(*pmd))) { 625 spin_unlock(&mm->page_table_lock); 626 mem_cgroup_uncharge_page(page); 627 put_page(page); 628 pte_free(mm, pgtable); 629 } else { 630 pmd_t entry; 631 entry = mk_pmd(page, vma->vm_page_prot); 632 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 633 entry = pmd_mkhuge(entry); 634 /* 635 * The spinlocking to take the lru_lock inside 636 * page_add_new_anon_rmap() acts as a full memory 637 * barrier to be sure clear_huge_page writes become 638 * visible after the set_pmd_at() write. 639 */ 640 page_add_new_anon_rmap(page, vma, haddr); 641 set_pmd_at(mm, haddr, pmd, entry); 642 pgtable_trans_huge_deposit(mm, pgtable); 643 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR); 644 mm->nr_ptes++; 645 spin_unlock(&mm->page_table_lock); 646 } 647 648 return 0; 649 } 650 651 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp) 652 { 653 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp; 654 } 655 656 static inline struct page *alloc_hugepage_vma(int defrag, 657 struct vm_area_struct *vma, 658 unsigned long haddr, int nd, 659 gfp_t extra_gfp) 660 { 661 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp), 662 HPAGE_PMD_ORDER, vma, haddr, nd); 663 } 664 665 #ifndef CONFIG_NUMA 666 static inline struct page *alloc_hugepage(int defrag) 667 { 668 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0), 669 HPAGE_PMD_ORDER); 670 } 671 #endif 672 673 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma, 674 unsigned long address, pmd_t *pmd, 675 unsigned int flags) 676 { 677 struct page *page; 678 unsigned long haddr = address & HPAGE_PMD_MASK; 679 pte_t *pte; 680 681 if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) { 682 if (unlikely(anon_vma_prepare(vma))) 683 return VM_FAULT_OOM; 684 if (unlikely(khugepaged_enter(vma))) 685 return VM_FAULT_OOM; 686 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma), 687 vma, haddr, numa_node_id(), 0); 688 if (unlikely(!page)) { 689 count_vm_event(THP_FAULT_FALLBACK); 690 goto out; 691 } 692 count_vm_event(THP_FAULT_ALLOC); 693 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) { 694 put_page(page); 695 goto out; 696 } 697 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, 698 page))) { 699 mem_cgroup_uncharge_page(page); 700 put_page(page); 701 goto out; 702 } 703 704 return 0; 705 } 706 out: 707 /* 708 * Use __pte_alloc instead of pte_alloc_map, because we can't 709 * run pte_offset_map on the pmd, if an huge pmd could 710 * materialize from under us from a different thread. 711 */ 712 if (unlikely(__pte_alloc(mm, vma, pmd, address))) 713 return VM_FAULT_OOM; 714 /* if an huge pmd materialized from under us just retry later */ 715 if (unlikely(pmd_trans_huge(*pmd))) 716 return 0; 717 /* 718 * A regular pmd is established and it can't morph into a huge pmd 719 * from under us anymore at this point because we hold the mmap_sem 720 * read mode and khugepaged takes it in write mode. So now it's 721 * safe to run pte_offset_map(). 722 */ 723 pte = pte_offset_map(pmd, address); 724 return handle_pte_fault(mm, vma, address, pte, pmd, flags); 725 } 726 727 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, 728 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, 729 struct vm_area_struct *vma) 730 { 731 struct page *src_page; 732 pmd_t pmd; 733 pgtable_t pgtable; 734 int ret; 735 736 ret = -ENOMEM; 737 pgtable = pte_alloc_one(dst_mm, addr); 738 if (unlikely(!pgtable)) 739 goto out; 740 741 spin_lock(&dst_mm->page_table_lock); 742 spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING); 743 744 ret = -EAGAIN; 745 pmd = *src_pmd; 746 if (unlikely(!pmd_trans_huge(pmd))) { 747 pte_free(dst_mm, pgtable); 748 goto out_unlock; 749 } 750 if (unlikely(pmd_trans_splitting(pmd))) { 751 /* split huge page running from under us */ 752 spin_unlock(&src_mm->page_table_lock); 753 spin_unlock(&dst_mm->page_table_lock); 754 pte_free(dst_mm, pgtable); 755 756 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */ 757 goto out; 758 } 759 src_page = pmd_page(pmd); 760 VM_BUG_ON(!PageHead(src_page)); 761 get_page(src_page); 762 page_dup_rmap(src_page); 763 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 764 765 pmdp_set_wrprotect(src_mm, addr, src_pmd); 766 pmd = pmd_mkold(pmd_wrprotect(pmd)); 767 set_pmd_at(dst_mm, addr, dst_pmd, pmd); 768 pgtable_trans_huge_deposit(dst_mm, pgtable); 769 dst_mm->nr_ptes++; 770 771 ret = 0; 772 out_unlock: 773 spin_unlock(&src_mm->page_table_lock); 774 spin_unlock(&dst_mm->page_table_lock); 775 out: 776 return ret; 777 } 778 779 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm, 780 struct vm_area_struct *vma, 781 unsigned long address, 782 pmd_t *pmd, pmd_t orig_pmd, 783 struct page *page, 784 unsigned long haddr) 785 { 786 pgtable_t pgtable; 787 pmd_t _pmd; 788 int ret = 0, i; 789 struct page **pages; 790 unsigned long mmun_start; /* For mmu_notifiers */ 791 unsigned long mmun_end; /* For mmu_notifiers */ 792 793 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR, 794 GFP_KERNEL); 795 if (unlikely(!pages)) { 796 ret |= VM_FAULT_OOM; 797 goto out; 798 } 799 800 for (i = 0; i < HPAGE_PMD_NR; i++) { 801 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE | 802 __GFP_OTHER_NODE, 803 vma, address, page_to_nid(page)); 804 if (unlikely(!pages[i] || 805 mem_cgroup_newpage_charge(pages[i], mm, 806 GFP_KERNEL))) { 807 if (pages[i]) 808 put_page(pages[i]); 809 mem_cgroup_uncharge_start(); 810 while (--i >= 0) { 811 mem_cgroup_uncharge_page(pages[i]); 812 put_page(pages[i]); 813 } 814 mem_cgroup_uncharge_end(); 815 kfree(pages); 816 ret |= VM_FAULT_OOM; 817 goto out; 818 } 819 } 820 821 for (i = 0; i < HPAGE_PMD_NR; i++) { 822 copy_user_highpage(pages[i], page + i, 823 haddr + PAGE_SIZE * i, vma); 824 __SetPageUptodate(pages[i]); 825 cond_resched(); 826 } 827 828 mmun_start = haddr; 829 mmun_end = haddr + HPAGE_PMD_SIZE; 830 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 831 832 spin_lock(&mm->page_table_lock); 833 if (unlikely(!pmd_same(*pmd, orig_pmd))) 834 goto out_free_pages; 835 VM_BUG_ON(!PageHead(page)); 836 837 pmdp_clear_flush(vma, haddr, pmd); 838 /* leave pmd empty until pte is filled */ 839 840 pgtable = pgtable_trans_huge_withdraw(mm); 841 pmd_populate(mm, &_pmd, pgtable); 842 843 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 844 pte_t *pte, entry; 845 entry = mk_pte(pages[i], vma->vm_page_prot); 846 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 847 page_add_new_anon_rmap(pages[i], vma, haddr); 848 pte = pte_offset_map(&_pmd, haddr); 849 VM_BUG_ON(!pte_none(*pte)); 850 set_pte_at(mm, haddr, pte, entry); 851 pte_unmap(pte); 852 } 853 kfree(pages); 854 855 smp_wmb(); /* make pte visible before pmd */ 856 pmd_populate(mm, pmd, pgtable); 857 page_remove_rmap(page); 858 spin_unlock(&mm->page_table_lock); 859 860 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 861 862 ret |= VM_FAULT_WRITE; 863 put_page(page); 864 865 out: 866 return ret; 867 868 out_free_pages: 869 spin_unlock(&mm->page_table_lock); 870 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 871 mem_cgroup_uncharge_start(); 872 for (i = 0; i < HPAGE_PMD_NR; i++) { 873 mem_cgroup_uncharge_page(pages[i]); 874 put_page(pages[i]); 875 } 876 mem_cgroup_uncharge_end(); 877 kfree(pages); 878 goto out; 879 } 880 881 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma, 882 unsigned long address, pmd_t *pmd, pmd_t orig_pmd) 883 { 884 int ret = 0; 885 struct page *page, *new_page; 886 unsigned long haddr; 887 unsigned long mmun_start; /* For mmu_notifiers */ 888 unsigned long mmun_end; /* For mmu_notifiers */ 889 890 VM_BUG_ON(!vma->anon_vma); 891 spin_lock(&mm->page_table_lock); 892 if (unlikely(!pmd_same(*pmd, orig_pmd))) 893 goto out_unlock; 894 895 page = pmd_page(orig_pmd); 896 VM_BUG_ON(!PageCompound(page) || !PageHead(page)); 897 haddr = address & HPAGE_PMD_MASK; 898 if (page_mapcount(page) == 1) { 899 pmd_t entry; 900 entry = pmd_mkyoung(orig_pmd); 901 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 902 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1)) 903 update_mmu_cache_pmd(vma, address, pmd); 904 ret |= VM_FAULT_WRITE; 905 goto out_unlock; 906 } 907 get_page(page); 908 spin_unlock(&mm->page_table_lock); 909 910 if (transparent_hugepage_enabled(vma) && 911 !transparent_hugepage_debug_cow()) 912 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma), 913 vma, haddr, numa_node_id(), 0); 914 else 915 new_page = NULL; 916 917 if (unlikely(!new_page)) { 918 count_vm_event(THP_FAULT_FALLBACK); 919 ret = do_huge_pmd_wp_page_fallback(mm, vma, address, 920 pmd, orig_pmd, page, haddr); 921 if (ret & VM_FAULT_OOM) 922 split_huge_page(page); 923 put_page(page); 924 goto out; 925 } 926 count_vm_event(THP_FAULT_ALLOC); 927 928 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) { 929 put_page(new_page); 930 split_huge_page(page); 931 put_page(page); 932 ret |= VM_FAULT_OOM; 933 goto out; 934 } 935 936 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR); 937 __SetPageUptodate(new_page); 938 939 mmun_start = haddr; 940 mmun_end = haddr + HPAGE_PMD_SIZE; 941 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 942 943 spin_lock(&mm->page_table_lock); 944 put_page(page); 945 if (unlikely(!pmd_same(*pmd, orig_pmd))) { 946 spin_unlock(&mm->page_table_lock); 947 mem_cgroup_uncharge_page(new_page); 948 put_page(new_page); 949 goto out_mn; 950 } else { 951 pmd_t entry; 952 VM_BUG_ON(!PageHead(page)); 953 entry = mk_pmd(new_page, vma->vm_page_prot); 954 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 955 entry = pmd_mkhuge(entry); 956 pmdp_clear_flush(vma, haddr, pmd); 957 page_add_new_anon_rmap(new_page, vma, haddr); 958 set_pmd_at(mm, haddr, pmd, entry); 959 update_mmu_cache_pmd(vma, address, pmd); 960 page_remove_rmap(page); 961 put_page(page); 962 ret |= VM_FAULT_WRITE; 963 } 964 spin_unlock(&mm->page_table_lock); 965 out_mn: 966 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 967 out: 968 return ret; 969 out_unlock: 970 spin_unlock(&mm->page_table_lock); 971 return ret; 972 } 973 974 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, 975 unsigned long addr, 976 pmd_t *pmd, 977 unsigned int flags) 978 { 979 struct mm_struct *mm = vma->vm_mm; 980 struct page *page = NULL; 981 982 assert_spin_locked(&mm->page_table_lock); 983 984 if (flags & FOLL_WRITE && !pmd_write(*pmd)) 985 goto out; 986 987 page = pmd_page(*pmd); 988 VM_BUG_ON(!PageHead(page)); 989 if (flags & FOLL_TOUCH) { 990 pmd_t _pmd; 991 /* 992 * We should set the dirty bit only for FOLL_WRITE but 993 * for now the dirty bit in the pmd is meaningless. 994 * And if the dirty bit will become meaningful and 995 * we'll only set it with FOLL_WRITE, an atomic 996 * set_bit will be required on the pmd to set the 997 * young bit, instead of the current set_pmd_at. 998 */ 999 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd)); 1000 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd); 1001 } 1002 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) { 1003 if (page->mapping && trylock_page(page)) { 1004 lru_add_drain(); 1005 if (page->mapping) 1006 mlock_vma_page(page); 1007 unlock_page(page); 1008 } 1009 } 1010 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; 1011 VM_BUG_ON(!PageCompound(page)); 1012 if (flags & FOLL_GET) 1013 get_page_foll(page); 1014 1015 out: 1016 return page; 1017 } 1018 1019 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1020 pmd_t *pmd, unsigned long addr) 1021 { 1022 int ret = 0; 1023 1024 if (__pmd_trans_huge_lock(pmd, vma) == 1) { 1025 struct page *page; 1026 pgtable_t pgtable; 1027 pmd_t orig_pmd; 1028 pgtable = pgtable_trans_huge_withdraw(tlb->mm); 1029 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd); 1030 page = pmd_page(orig_pmd); 1031 tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 1032 page_remove_rmap(page); 1033 VM_BUG_ON(page_mapcount(page) < 0); 1034 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); 1035 VM_BUG_ON(!PageHead(page)); 1036 tlb->mm->nr_ptes--; 1037 spin_unlock(&tlb->mm->page_table_lock); 1038 tlb_remove_page(tlb, page); 1039 pte_free(tlb->mm, pgtable); 1040 ret = 1; 1041 } 1042 return ret; 1043 } 1044 1045 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 1046 unsigned long addr, unsigned long end, 1047 unsigned char *vec) 1048 { 1049 int ret = 0; 1050 1051 if (__pmd_trans_huge_lock(pmd, vma) == 1) { 1052 /* 1053 * All logical pages in the range are present 1054 * if backed by a huge page. 1055 */ 1056 spin_unlock(&vma->vm_mm->page_table_lock); 1057 memset(vec, 1, (end - addr) >> PAGE_SHIFT); 1058 ret = 1; 1059 } 1060 1061 return ret; 1062 } 1063 1064 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma, 1065 unsigned long old_addr, 1066 unsigned long new_addr, unsigned long old_end, 1067 pmd_t *old_pmd, pmd_t *new_pmd) 1068 { 1069 int ret = 0; 1070 pmd_t pmd; 1071 1072 struct mm_struct *mm = vma->vm_mm; 1073 1074 if ((old_addr & ~HPAGE_PMD_MASK) || 1075 (new_addr & ~HPAGE_PMD_MASK) || 1076 old_end - old_addr < HPAGE_PMD_SIZE || 1077 (new_vma->vm_flags & VM_NOHUGEPAGE)) 1078 goto out; 1079 1080 /* 1081 * The destination pmd shouldn't be established, free_pgtables() 1082 * should have release it. 1083 */ 1084 if (WARN_ON(!pmd_none(*new_pmd))) { 1085 VM_BUG_ON(pmd_trans_huge(*new_pmd)); 1086 goto out; 1087 } 1088 1089 ret = __pmd_trans_huge_lock(old_pmd, vma); 1090 if (ret == 1) { 1091 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd); 1092 VM_BUG_ON(!pmd_none(*new_pmd)); 1093 set_pmd_at(mm, new_addr, new_pmd, pmd); 1094 spin_unlock(&mm->page_table_lock); 1095 } 1096 out: 1097 return ret; 1098 } 1099 1100 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 1101 unsigned long addr, pgprot_t newprot) 1102 { 1103 struct mm_struct *mm = vma->vm_mm; 1104 int ret = 0; 1105 1106 if (__pmd_trans_huge_lock(pmd, vma) == 1) { 1107 pmd_t entry; 1108 entry = pmdp_get_and_clear(mm, addr, pmd); 1109 entry = pmd_modify(entry, newprot); 1110 set_pmd_at(mm, addr, pmd, entry); 1111 spin_unlock(&vma->vm_mm->page_table_lock); 1112 ret = 1; 1113 } 1114 1115 return ret; 1116 } 1117 1118 /* 1119 * Returns 1 if a given pmd maps a stable (not under splitting) thp. 1120 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise. 1121 * 1122 * Note that if it returns 1, this routine returns without unlocking page 1123 * table locks. So callers must unlock them. 1124 */ 1125 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma) 1126 { 1127 spin_lock(&vma->vm_mm->page_table_lock); 1128 if (likely(pmd_trans_huge(*pmd))) { 1129 if (unlikely(pmd_trans_splitting(*pmd))) { 1130 spin_unlock(&vma->vm_mm->page_table_lock); 1131 wait_split_huge_page(vma->anon_vma, pmd); 1132 return -1; 1133 } else { 1134 /* Thp mapped by 'pmd' is stable, so we can 1135 * handle it as it is. */ 1136 return 1; 1137 } 1138 } 1139 spin_unlock(&vma->vm_mm->page_table_lock); 1140 return 0; 1141 } 1142 1143 pmd_t *page_check_address_pmd(struct page *page, 1144 struct mm_struct *mm, 1145 unsigned long address, 1146 enum page_check_address_pmd_flag flag) 1147 { 1148 pgd_t *pgd; 1149 pud_t *pud; 1150 pmd_t *pmd, *ret = NULL; 1151 1152 if (address & ~HPAGE_PMD_MASK) 1153 goto out; 1154 1155 pgd = pgd_offset(mm, address); 1156 if (!pgd_present(*pgd)) 1157 goto out; 1158 1159 pud = pud_offset(pgd, address); 1160 if (!pud_present(*pud)) 1161 goto out; 1162 1163 pmd = pmd_offset(pud, address); 1164 if (pmd_none(*pmd)) 1165 goto out; 1166 if (pmd_page(*pmd) != page) 1167 goto out; 1168 /* 1169 * split_vma() may create temporary aliased mappings. There is 1170 * no risk as long as all huge pmd are found and have their 1171 * splitting bit set before __split_huge_page_refcount 1172 * runs. Finding the same huge pmd more than once during the 1173 * same rmap walk is not a problem. 1174 */ 1175 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG && 1176 pmd_trans_splitting(*pmd)) 1177 goto out; 1178 if (pmd_trans_huge(*pmd)) { 1179 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG && 1180 !pmd_trans_splitting(*pmd)); 1181 ret = pmd; 1182 } 1183 out: 1184 return ret; 1185 } 1186 1187 static int __split_huge_page_splitting(struct page *page, 1188 struct vm_area_struct *vma, 1189 unsigned long address) 1190 { 1191 struct mm_struct *mm = vma->vm_mm; 1192 pmd_t *pmd; 1193 int ret = 0; 1194 /* For mmu_notifiers */ 1195 const unsigned long mmun_start = address; 1196 const unsigned long mmun_end = address + HPAGE_PMD_SIZE; 1197 1198 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 1199 spin_lock(&mm->page_table_lock); 1200 pmd = page_check_address_pmd(page, mm, address, 1201 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG); 1202 if (pmd) { 1203 /* 1204 * We can't temporarily set the pmd to null in order 1205 * to split it, the pmd must remain marked huge at all 1206 * times or the VM won't take the pmd_trans_huge paths 1207 * and it won't wait on the anon_vma->root->mutex to 1208 * serialize against split_huge_page*. 1209 */ 1210 pmdp_splitting_flush(vma, address, pmd); 1211 ret = 1; 1212 } 1213 spin_unlock(&mm->page_table_lock); 1214 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 1215 1216 return ret; 1217 } 1218 1219 static void __split_huge_page_refcount(struct page *page) 1220 { 1221 int i; 1222 struct zone *zone = page_zone(page); 1223 struct lruvec *lruvec; 1224 int tail_count = 0; 1225 1226 /* prevent PageLRU to go away from under us, and freeze lru stats */ 1227 spin_lock_irq(&zone->lru_lock); 1228 lruvec = mem_cgroup_page_lruvec(page, zone); 1229 1230 compound_lock(page); 1231 /* complete memcg works before add pages to LRU */ 1232 mem_cgroup_split_huge_fixup(page); 1233 1234 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) { 1235 struct page *page_tail = page + i; 1236 1237 /* tail_page->_mapcount cannot change */ 1238 BUG_ON(page_mapcount(page_tail) < 0); 1239 tail_count += page_mapcount(page_tail); 1240 /* check for overflow */ 1241 BUG_ON(tail_count < 0); 1242 BUG_ON(atomic_read(&page_tail->_count) != 0); 1243 /* 1244 * tail_page->_count is zero and not changing from 1245 * under us. But get_page_unless_zero() may be running 1246 * from under us on the tail_page. If we used 1247 * atomic_set() below instead of atomic_add(), we 1248 * would then run atomic_set() concurrently with 1249 * get_page_unless_zero(), and atomic_set() is 1250 * implemented in C not using locked ops. spin_unlock 1251 * on x86 sometime uses locked ops because of PPro 1252 * errata 66, 92, so unless somebody can guarantee 1253 * atomic_set() here would be safe on all archs (and 1254 * not only on x86), it's safer to use atomic_add(). 1255 */ 1256 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1, 1257 &page_tail->_count); 1258 1259 /* after clearing PageTail the gup refcount can be released */ 1260 smp_mb(); 1261 1262 /* 1263 * retain hwpoison flag of the poisoned tail page: 1264 * fix for the unsuitable process killed on Guest Machine(KVM) 1265 * by the memory-failure. 1266 */ 1267 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON; 1268 page_tail->flags |= (page->flags & 1269 ((1L << PG_referenced) | 1270 (1L << PG_swapbacked) | 1271 (1L << PG_mlocked) | 1272 (1L << PG_uptodate))); 1273 page_tail->flags |= (1L << PG_dirty); 1274 1275 /* clear PageTail before overwriting first_page */ 1276 smp_wmb(); 1277 1278 /* 1279 * __split_huge_page_splitting() already set the 1280 * splitting bit in all pmd that could map this 1281 * hugepage, that will ensure no CPU can alter the 1282 * mapcount on the head page. The mapcount is only 1283 * accounted in the head page and it has to be 1284 * transferred to all tail pages in the below code. So 1285 * for this code to be safe, the split the mapcount 1286 * can't change. But that doesn't mean userland can't 1287 * keep changing and reading the page contents while 1288 * we transfer the mapcount, so the pmd splitting 1289 * status is achieved setting a reserved bit in the 1290 * pmd, not by clearing the present bit. 1291 */ 1292 page_tail->_mapcount = page->_mapcount; 1293 1294 BUG_ON(page_tail->mapping); 1295 page_tail->mapping = page->mapping; 1296 1297 page_tail->index = page->index + i; 1298 1299 BUG_ON(!PageAnon(page_tail)); 1300 BUG_ON(!PageUptodate(page_tail)); 1301 BUG_ON(!PageDirty(page_tail)); 1302 BUG_ON(!PageSwapBacked(page_tail)); 1303 1304 lru_add_page_tail(page, page_tail, lruvec); 1305 } 1306 atomic_sub(tail_count, &page->_count); 1307 BUG_ON(atomic_read(&page->_count) <= 0); 1308 1309 __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1); 1310 __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR); 1311 1312 ClearPageCompound(page); 1313 compound_unlock(page); 1314 spin_unlock_irq(&zone->lru_lock); 1315 1316 for (i = 1; i < HPAGE_PMD_NR; i++) { 1317 struct page *page_tail = page + i; 1318 BUG_ON(page_count(page_tail) <= 0); 1319 /* 1320 * Tail pages may be freed if there wasn't any mapping 1321 * like if add_to_swap() is running on a lru page that 1322 * had its mapping zapped. And freeing these pages 1323 * requires taking the lru_lock so we do the put_page 1324 * of the tail pages after the split is complete. 1325 */ 1326 put_page(page_tail); 1327 } 1328 1329 /* 1330 * Only the head page (now become a regular page) is required 1331 * to be pinned by the caller. 1332 */ 1333 BUG_ON(page_count(page) <= 0); 1334 } 1335 1336 static int __split_huge_page_map(struct page *page, 1337 struct vm_area_struct *vma, 1338 unsigned long address) 1339 { 1340 struct mm_struct *mm = vma->vm_mm; 1341 pmd_t *pmd, _pmd; 1342 int ret = 0, i; 1343 pgtable_t pgtable; 1344 unsigned long haddr; 1345 1346 spin_lock(&mm->page_table_lock); 1347 pmd = page_check_address_pmd(page, mm, address, 1348 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG); 1349 if (pmd) { 1350 pgtable = pgtable_trans_huge_withdraw(mm); 1351 pmd_populate(mm, &_pmd, pgtable); 1352 1353 haddr = address; 1354 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 1355 pte_t *pte, entry; 1356 BUG_ON(PageCompound(page+i)); 1357 entry = mk_pte(page + i, vma->vm_page_prot); 1358 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 1359 if (!pmd_write(*pmd)) 1360 entry = pte_wrprotect(entry); 1361 else 1362 BUG_ON(page_mapcount(page) != 1); 1363 if (!pmd_young(*pmd)) 1364 entry = pte_mkold(entry); 1365 pte = pte_offset_map(&_pmd, haddr); 1366 BUG_ON(!pte_none(*pte)); 1367 set_pte_at(mm, haddr, pte, entry); 1368 pte_unmap(pte); 1369 } 1370 1371 smp_wmb(); /* make pte visible before pmd */ 1372 /* 1373 * Up to this point the pmd is present and huge and 1374 * userland has the whole access to the hugepage 1375 * during the split (which happens in place). If we 1376 * overwrite the pmd with the not-huge version 1377 * pointing to the pte here (which of course we could 1378 * if all CPUs were bug free), userland could trigger 1379 * a small page size TLB miss on the small sized TLB 1380 * while the hugepage TLB entry is still established 1381 * in the huge TLB. Some CPU doesn't like that. See 1382 * http://support.amd.com/us/Processor_TechDocs/41322.pdf, 1383 * Erratum 383 on page 93. Intel should be safe but is 1384 * also warns that it's only safe if the permission 1385 * and cache attributes of the two entries loaded in 1386 * the two TLB is identical (which should be the case 1387 * here). But it is generally safer to never allow 1388 * small and huge TLB entries for the same virtual 1389 * address to be loaded simultaneously. So instead of 1390 * doing "pmd_populate(); flush_tlb_range();" we first 1391 * mark the current pmd notpresent (atomically because 1392 * here the pmd_trans_huge and pmd_trans_splitting 1393 * must remain set at all times on the pmd until the 1394 * split is complete for this pmd), then we flush the 1395 * SMP TLB and finally we write the non-huge version 1396 * of the pmd entry with pmd_populate. 1397 */ 1398 pmdp_invalidate(vma, address, pmd); 1399 pmd_populate(mm, pmd, pgtable); 1400 ret = 1; 1401 } 1402 spin_unlock(&mm->page_table_lock); 1403 1404 return ret; 1405 } 1406 1407 /* must be called with anon_vma->root->mutex hold */ 1408 static void __split_huge_page(struct page *page, 1409 struct anon_vma *anon_vma) 1410 { 1411 int mapcount, mapcount2; 1412 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); 1413 struct anon_vma_chain *avc; 1414 1415 BUG_ON(!PageHead(page)); 1416 BUG_ON(PageTail(page)); 1417 1418 mapcount = 0; 1419 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) { 1420 struct vm_area_struct *vma = avc->vma; 1421 unsigned long addr = vma_address(page, vma); 1422 BUG_ON(is_vma_temporary_stack(vma)); 1423 mapcount += __split_huge_page_splitting(page, vma, addr); 1424 } 1425 /* 1426 * It is critical that new vmas are added to the tail of the 1427 * anon_vma list. This guarantes that if copy_huge_pmd() runs 1428 * and establishes a child pmd before 1429 * __split_huge_page_splitting() freezes the parent pmd (so if 1430 * we fail to prevent copy_huge_pmd() from running until the 1431 * whole __split_huge_page() is complete), we will still see 1432 * the newly established pmd of the child later during the 1433 * walk, to be able to set it as pmd_trans_splitting too. 1434 */ 1435 if (mapcount != page_mapcount(page)) 1436 printk(KERN_ERR "mapcount %d page_mapcount %d\n", 1437 mapcount, page_mapcount(page)); 1438 BUG_ON(mapcount != page_mapcount(page)); 1439 1440 __split_huge_page_refcount(page); 1441 1442 mapcount2 = 0; 1443 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) { 1444 struct vm_area_struct *vma = avc->vma; 1445 unsigned long addr = vma_address(page, vma); 1446 BUG_ON(is_vma_temporary_stack(vma)); 1447 mapcount2 += __split_huge_page_map(page, vma, addr); 1448 } 1449 if (mapcount != mapcount2) 1450 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n", 1451 mapcount, mapcount2, page_mapcount(page)); 1452 BUG_ON(mapcount != mapcount2); 1453 } 1454 1455 int split_huge_page(struct page *page) 1456 { 1457 struct anon_vma *anon_vma; 1458 int ret = 1; 1459 1460 BUG_ON(!PageAnon(page)); 1461 anon_vma = page_lock_anon_vma(page); 1462 if (!anon_vma) 1463 goto out; 1464 ret = 0; 1465 if (!PageCompound(page)) 1466 goto out_unlock; 1467 1468 BUG_ON(!PageSwapBacked(page)); 1469 __split_huge_page(page, anon_vma); 1470 count_vm_event(THP_SPLIT); 1471 1472 BUG_ON(PageCompound(page)); 1473 out_unlock: 1474 page_unlock_anon_vma(anon_vma); 1475 out: 1476 return ret; 1477 } 1478 1479 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE) 1480 1481 int hugepage_madvise(struct vm_area_struct *vma, 1482 unsigned long *vm_flags, int advice) 1483 { 1484 struct mm_struct *mm = vma->vm_mm; 1485 1486 switch (advice) { 1487 case MADV_HUGEPAGE: 1488 /* 1489 * Be somewhat over-protective like KSM for now! 1490 */ 1491 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP)) 1492 return -EINVAL; 1493 if (mm->def_flags & VM_NOHUGEPAGE) 1494 return -EINVAL; 1495 *vm_flags &= ~VM_NOHUGEPAGE; 1496 *vm_flags |= VM_HUGEPAGE; 1497 /* 1498 * If the vma become good for khugepaged to scan, 1499 * register it here without waiting a page fault that 1500 * may not happen any time soon. 1501 */ 1502 if (unlikely(khugepaged_enter_vma_merge(vma))) 1503 return -ENOMEM; 1504 break; 1505 case MADV_NOHUGEPAGE: 1506 /* 1507 * Be somewhat over-protective like KSM for now! 1508 */ 1509 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP)) 1510 return -EINVAL; 1511 *vm_flags &= ~VM_HUGEPAGE; 1512 *vm_flags |= VM_NOHUGEPAGE; 1513 /* 1514 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning 1515 * this vma even if we leave the mm registered in khugepaged if 1516 * it got registered before VM_NOHUGEPAGE was set. 1517 */ 1518 break; 1519 } 1520 1521 return 0; 1522 } 1523 1524 static int __init khugepaged_slab_init(void) 1525 { 1526 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot", 1527 sizeof(struct mm_slot), 1528 __alignof__(struct mm_slot), 0, NULL); 1529 if (!mm_slot_cache) 1530 return -ENOMEM; 1531 1532 return 0; 1533 } 1534 1535 static void __init khugepaged_slab_free(void) 1536 { 1537 kmem_cache_destroy(mm_slot_cache); 1538 mm_slot_cache = NULL; 1539 } 1540 1541 static inline struct mm_slot *alloc_mm_slot(void) 1542 { 1543 if (!mm_slot_cache) /* initialization failed */ 1544 return NULL; 1545 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL); 1546 } 1547 1548 static inline void free_mm_slot(struct mm_slot *mm_slot) 1549 { 1550 kmem_cache_free(mm_slot_cache, mm_slot); 1551 } 1552 1553 static int __init mm_slots_hash_init(void) 1554 { 1555 mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head), 1556 GFP_KERNEL); 1557 if (!mm_slots_hash) 1558 return -ENOMEM; 1559 return 0; 1560 } 1561 1562 #if 0 1563 static void __init mm_slots_hash_free(void) 1564 { 1565 kfree(mm_slots_hash); 1566 mm_slots_hash = NULL; 1567 } 1568 #endif 1569 1570 static struct mm_slot *get_mm_slot(struct mm_struct *mm) 1571 { 1572 struct mm_slot *mm_slot; 1573 struct hlist_head *bucket; 1574 struct hlist_node *node; 1575 1576 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct)) 1577 % MM_SLOTS_HASH_HEADS]; 1578 hlist_for_each_entry(mm_slot, node, bucket, hash) { 1579 if (mm == mm_slot->mm) 1580 return mm_slot; 1581 } 1582 return NULL; 1583 } 1584 1585 static void insert_to_mm_slots_hash(struct mm_struct *mm, 1586 struct mm_slot *mm_slot) 1587 { 1588 struct hlist_head *bucket; 1589 1590 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct)) 1591 % MM_SLOTS_HASH_HEADS]; 1592 mm_slot->mm = mm; 1593 hlist_add_head(&mm_slot->hash, bucket); 1594 } 1595 1596 static inline int khugepaged_test_exit(struct mm_struct *mm) 1597 { 1598 return atomic_read(&mm->mm_users) == 0; 1599 } 1600 1601 int __khugepaged_enter(struct mm_struct *mm) 1602 { 1603 struct mm_slot *mm_slot; 1604 int wakeup; 1605 1606 mm_slot = alloc_mm_slot(); 1607 if (!mm_slot) 1608 return -ENOMEM; 1609 1610 /* __khugepaged_exit() must not run from under us */ 1611 VM_BUG_ON(khugepaged_test_exit(mm)); 1612 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) { 1613 free_mm_slot(mm_slot); 1614 return 0; 1615 } 1616 1617 spin_lock(&khugepaged_mm_lock); 1618 insert_to_mm_slots_hash(mm, mm_slot); 1619 /* 1620 * Insert just behind the scanning cursor, to let the area settle 1621 * down a little. 1622 */ 1623 wakeup = list_empty(&khugepaged_scan.mm_head); 1624 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head); 1625 spin_unlock(&khugepaged_mm_lock); 1626 1627 atomic_inc(&mm->mm_count); 1628 if (wakeup) 1629 wake_up_interruptible(&khugepaged_wait); 1630 1631 return 0; 1632 } 1633 1634 int khugepaged_enter_vma_merge(struct vm_area_struct *vma) 1635 { 1636 unsigned long hstart, hend; 1637 if (!vma->anon_vma) 1638 /* 1639 * Not yet faulted in so we will register later in the 1640 * page fault if needed. 1641 */ 1642 return 0; 1643 if (vma->vm_ops) 1644 /* khugepaged not yet working on file or special mappings */ 1645 return 0; 1646 VM_BUG_ON(vma->vm_flags & VM_NO_THP); 1647 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; 1648 hend = vma->vm_end & HPAGE_PMD_MASK; 1649 if (hstart < hend) 1650 return khugepaged_enter(vma); 1651 return 0; 1652 } 1653 1654 void __khugepaged_exit(struct mm_struct *mm) 1655 { 1656 struct mm_slot *mm_slot; 1657 int free = 0; 1658 1659 spin_lock(&khugepaged_mm_lock); 1660 mm_slot = get_mm_slot(mm); 1661 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) { 1662 hlist_del(&mm_slot->hash); 1663 list_del(&mm_slot->mm_node); 1664 free = 1; 1665 } 1666 spin_unlock(&khugepaged_mm_lock); 1667 1668 if (free) { 1669 clear_bit(MMF_VM_HUGEPAGE, &mm->flags); 1670 free_mm_slot(mm_slot); 1671 mmdrop(mm); 1672 } else if (mm_slot) { 1673 /* 1674 * This is required to serialize against 1675 * khugepaged_test_exit() (which is guaranteed to run 1676 * under mmap sem read mode). Stop here (after we 1677 * return all pagetables will be destroyed) until 1678 * khugepaged has finished working on the pagetables 1679 * under the mmap_sem. 1680 */ 1681 down_write(&mm->mmap_sem); 1682 up_write(&mm->mmap_sem); 1683 } 1684 } 1685 1686 static void release_pte_page(struct page *page) 1687 { 1688 /* 0 stands for page_is_file_cache(page) == false */ 1689 dec_zone_page_state(page, NR_ISOLATED_ANON + 0); 1690 unlock_page(page); 1691 putback_lru_page(page); 1692 } 1693 1694 static void release_pte_pages(pte_t *pte, pte_t *_pte) 1695 { 1696 while (--_pte >= pte) { 1697 pte_t pteval = *_pte; 1698 if (!pte_none(pteval)) 1699 release_pte_page(pte_page(pteval)); 1700 } 1701 } 1702 1703 static void release_all_pte_pages(pte_t *pte) 1704 { 1705 release_pte_pages(pte, pte + HPAGE_PMD_NR); 1706 } 1707 1708 static int __collapse_huge_page_isolate(struct vm_area_struct *vma, 1709 unsigned long address, 1710 pte_t *pte) 1711 { 1712 struct page *page; 1713 pte_t *_pte; 1714 int referenced = 0, isolated = 0, none = 0; 1715 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; 1716 _pte++, address += PAGE_SIZE) { 1717 pte_t pteval = *_pte; 1718 if (pte_none(pteval)) { 1719 if (++none <= khugepaged_max_ptes_none) 1720 continue; 1721 else { 1722 release_pte_pages(pte, _pte); 1723 goto out; 1724 } 1725 } 1726 if (!pte_present(pteval) || !pte_write(pteval)) { 1727 release_pte_pages(pte, _pte); 1728 goto out; 1729 } 1730 page = vm_normal_page(vma, address, pteval); 1731 if (unlikely(!page)) { 1732 release_pte_pages(pte, _pte); 1733 goto out; 1734 } 1735 VM_BUG_ON(PageCompound(page)); 1736 BUG_ON(!PageAnon(page)); 1737 VM_BUG_ON(!PageSwapBacked(page)); 1738 1739 /* cannot use mapcount: can't collapse if there's a gup pin */ 1740 if (page_count(page) != 1) { 1741 release_pte_pages(pte, _pte); 1742 goto out; 1743 } 1744 /* 1745 * We can do it before isolate_lru_page because the 1746 * page can't be freed from under us. NOTE: PG_lock 1747 * is needed to serialize against split_huge_page 1748 * when invoked from the VM. 1749 */ 1750 if (!trylock_page(page)) { 1751 release_pte_pages(pte, _pte); 1752 goto out; 1753 } 1754 /* 1755 * Isolate the page to avoid collapsing an hugepage 1756 * currently in use by the VM. 1757 */ 1758 if (isolate_lru_page(page)) { 1759 unlock_page(page); 1760 release_pte_pages(pte, _pte); 1761 goto out; 1762 } 1763 /* 0 stands for page_is_file_cache(page) == false */ 1764 inc_zone_page_state(page, NR_ISOLATED_ANON + 0); 1765 VM_BUG_ON(!PageLocked(page)); 1766 VM_BUG_ON(PageLRU(page)); 1767 1768 /* If there is no mapped pte young don't collapse the page */ 1769 if (pte_young(pteval) || PageReferenced(page) || 1770 mmu_notifier_test_young(vma->vm_mm, address)) 1771 referenced = 1; 1772 } 1773 if (unlikely(!referenced)) 1774 release_all_pte_pages(pte); 1775 else 1776 isolated = 1; 1777 out: 1778 return isolated; 1779 } 1780 1781 static void __collapse_huge_page_copy(pte_t *pte, struct page *page, 1782 struct vm_area_struct *vma, 1783 unsigned long address, 1784 spinlock_t *ptl) 1785 { 1786 pte_t *_pte; 1787 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) { 1788 pte_t pteval = *_pte; 1789 struct page *src_page; 1790 1791 if (pte_none(pteval)) { 1792 clear_user_highpage(page, address); 1793 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); 1794 } else { 1795 src_page = pte_page(pteval); 1796 copy_user_highpage(page, src_page, address, vma); 1797 VM_BUG_ON(page_mapcount(src_page) != 1); 1798 release_pte_page(src_page); 1799 /* 1800 * ptl mostly unnecessary, but preempt has to 1801 * be disabled to update the per-cpu stats 1802 * inside page_remove_rmap(). 1803 */ 1804 spin_lock(ptl); 1805 /* 1806 * paravirt calls inside pte_clear here are 1807 * superfluous. 1808 */ 1809 pte_clear(vma->vm_mm, address, _pte); 1810 page_remove_rmap(src_page); 1811 spin_unlock(ptl); 1812 free_page_and_swap_cache(src_page); 1813 } 1814 1815 address += PAGE_SIZE; 1816 page++; 1817 } 1818 } 1819 1820 static void khugepaged_alloc_sleep(void) 1821 { 1822 wait_event_freezable_timeout(khugepaged_wait, false, 1823 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs)); 1824 } 1825 1826 #ifdef CONFIG_NUMA 1827 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) 1828 { 1829 if (IS_ERR(*hpage)) { 1830 if (!*wait) 1831 return false; 1832 1833 *wait = false; 1834 *hpage = NULL; 1835 khugepaged_alloc_sleep(); 1836 } else if (*hpage) { 1837 put_page(*hpage); 1838 *hpage = NULL; 1839 } 1840 1841 return true; 1842 } 1843 1844 static struct page 1845 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm, 1846 struct vm_area_struct *vma, unsigned long address, 1847 int node) 1848 { 1849 VM_BUG_ON(*hpage); 1850 /* 1851 * Allocate the page while the vma is still valid and under 1852 * the mmap_sem read mode so there is no memory allocation 1853 * later when we take the mmap_sem in write mode. This is more 1854 * friendly behavior (OTOH it may actually hide bugs) to 1855 * filesystems in userland with daemons allocating memory in 1856 * the userland I/O paths. Allocating memory with the 1857 * mmap_sem in read mode is good idea also to allow greater 1858 * scalability. 1859 */ 1860 *hpage = alloc_hugepage_vma(khugepaged_defrag(), vma, address, 1861 node, __GFP_OTHER_NODE); 1862 1863 /* 1864 * After allocating the hugepage, release the mmap_sem read lock in 1865 * preparation for taking it in write mode. 1866 */ 1867 up_read(&mm->mmap_sem); 1868 if (unlikely(!*hpage)) { 1869 count_vm_event(THP_COLLAPSE_ALLOC_FAILED); 1870 *hpage = ERR_PTR(-ENOMEM); 1871 return NULL; 1872 } 1873 1874 count_vm_event(THP_COLLAPSE_ALLOC); 1875 return *hpage; 1876 } 1877 #else 1878 static struct page *khugepaged_alloc_hugepage(bool *wait) 1879 { 1880 struct page *hpage; 1881 1882 do { 1883 hpage = alloc_hugepage(khugepaged_defrag()); 1884 if (!hpage) { 1885 count_vm_event(THP_COLLAPSE_ALLOC_FAILED); 1886 if (!*wait) 1887 return NULL; 1888 1889 *wait = false; 1890 khugepaged_alloc_sleep(); 1891 } else 1892 count_vm_event(THP_COLLAPSE_ALLOC); 1893 } while (unlikely(!hpage) && likely(khugepaged_enabled())); 1894 1895 return hpage; 1896 } 1897 1898 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) 1899 { 1900 if (!*hpage) 1901 *hpage = khugepaged_alloc_hugepage(wait); 1902 1903 if (unlikely(!*hpage)) 1904 return false; 1905 1906 return true; 1907 } 1908 1909 static struct page 1910 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm, 1911 struct vm_area_struct *vma, unsigned long address, 1912 int node) 1913 { 1914 up_read(&mm->mmap_sem); 1915 VM_BUG_ON(!*hpage); 1916 return *hpage; 1917 } 1918 #endif 1919 1920 static void collapse_huge_page(struct mm_struct *mm, 1921 unsigned long address, 1922 struct page **hpage, 1923 struct vm_area_struct *vma, 1924 int node) 1925 { 1926 pgd_t *pgd; 1927 pud_t *pud; 1928 pmd_t *pmd, _pmd; 1929 pte_t *pte; 1930 pgtable_t pgtable; 1931 struct page *new_page; 1932 spinlock_t *ptl; 1933 int isolated; 1934 unsigned long hstart, hend; 1935 unsigned long mmun_start; /* For mmu_notifiers */ 1936 unsigned long mmun_end; /* For mmu_notifiers */ 1937 1938 VM_BUG_ON(address & ~HPAGE_PMD_MASK); 1939 1940 /* release the mmap_sem read lock. */ 1941 new_page = khugepaged_alloc_page(hpage, mm, vma, address, node); 1942 if (!new_page) 1943 return; 1944 1945 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) 1946 return; 1947 1948 /* 1949 * Prevent all access to pagetables with the exception of 1950 * gup_fast later hanlded by the ptep_clear_flush and the VM 1951 * handled by the anon_vma lock + PG_lock. 1952 */ 1953 down_write(&mm->mmap_sem); 1954 if (unlikely(khugepaged_test_exit(mm))) 1955 goto out; 1956 1957 vma = find_vma(mm, address); 1958 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; 1959 hend = vma->vm_end & HPAGE_PMD_MASK; 1960 if (address < hstart || address + HPAGE_PMD_SIZE > hend) 1961 goto out; 1962 1963 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) || 1964 (vma->vm_flags & VM_NOHUGEPAGE)) 1965 goto out; 1966 1967 if (!vma->anon_vma || vma->vm_ops) 1968 goto out; 1969 if (is_vma_temporary_stack(vma)) 1970 goto out; 1971 VM_BUG_ON(vma->vm_flags & VM_NO_THP); 1972 1973 pgd = pgd_offset(mm, address); 1974 if (!pgd_present(*pgd)) 1975 goto out; 1976 1977 pud = pud_offset(pgd, address); 1978 if (!pud_present(*pud)) 1979 goto out; 1980 1981 pmd = pmd_offset(pud, address); 1982 /* pmd can't go away or become huge under us */ 1983 if (!pmd_present(*pmd) || pmd_trans_huge(*pmd)) 1984 goto out; 1985 1986 anon_vma_lock(vma->anon_vma); 1987 1988 pte = pte_offset_map(pmd, address); 1989 ptl = pte_lockptr(mm, pmd); 1990 1991 mmun_start = address; 1992 mmun_end = address + HPAGE_PMD_SIZE; 1993 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 1994 spin_lock(&mm->page_table_lock); /* probably unnecessary */ 1995 /* 1996 * After this gup_fast can't run anymore. This also removes 1997 * any huge TLB entry from the CPU so we won't allow 1998 * huge and small TLB entries for the same virtual address 1999 * to avoid the risk of CPU bugs in that area. 2000 */ 2001 _pmd = pmdp_clear_flush(vma, address, pmd); 2002 spin_unlock(&mm->page_table_lock); 2003 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 2004 2005 spin_lock(ptl); 2006 isolated = __collapse_huge_page_isolate(vma, address, pte); 2007 spin_unlock(ptl); 2008 2009 if (unlikely(!isolated)) { 2010 pte_unmap(pte); 2011 spin_lock(&mm->page_table_lock); 2012 BUG_ON(!pmd_none(*pmd)); 2013 set_pmd_at(mm, address, pmd, _pmd); 2014 spin_unlock(&mm->page_table_lock); 2015 anon_vma_unlock(vma->anon_vma); 2016 goto out; 2017 } 2018 2019 /* 2020 * All pages are isolated and locked so anon_vma rmap 2021 * can't run anymore. 2022 */ 2023 anon_vma_unlock(vma->anon_vma); 2024 2025 __collapse_huge_page_copy(pte, new_page, vma, address, ptl); 2026 pte_unmap(pte); 2027 __SetPageUptodate(new_page); 2028 pgtable = pmd_pgtable(_pmd); 2029 2030 _pmd = mk_pmd(new_page, vma->vm_page_prot); 2031 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma); 2032 _pmd = pmd_mkhuge(_pmd); 2033 2034 /* 2035 * spin_lock() below is not the equivalent of smp_wmb(), so 2036 * this is needed to avoid the copy_huge_page writes to become 2037 * visible after the set_pmd_at() write. 2038 */ 2039 smp_wmb(); 2040 2041 spin_lock(&mm->page_table_lock); 2042 BUG_ON(!pmd_none(*pmd)); 2043 page_add_new_anon_rmap(new_page, vma, address); 2044 set_pmd_at(mm, address, pmd, _pmd); 2045 update_mmu_cache_pmd(vma, address, pmd); 2046 pgtable_trans_huge_deposit(mm, pgtable); 2047 spin_unlock(&mm->page_table_lock); 2048 2049 *hpage = NULL; 2050 2051 khugepaged_pages_collapsed++; 2052 out_up_write: 2053 up_write(&mm->mmap_sem); 2054 return; 2055 2056 out: 2057 mem_cgroup_uncharge_page(new_page); 2058 goto out_up_write; 2059 } 2060 2061 static int khugepaged_scan_pmd(struct mm_struct *mm, 2062 struct vm_area_struct *vma, 2063 unsigned long address, 2064 struct page **hpage) 2065 { 2066 pgd_t *pgd; 2067 pud_t *pud; 2068 pmd_t *pmd; 2069 pte_t *pte, *_pte; 2070 int ret = 0, referenced = 0, none = 0; 2071 struct page *page; 2072 unsigned long _address; 2073 spinlock_t *ptl; 2074 int node = -1; 2075 2076 VM_BUG_ON(address & ~HPAGE_PMD_MASK); 2077 2078 pgd = pgd_offset(mm, address); 2079 if (!pgd_present(*pgd)) 2080 goto out; 2081 2082 pud = pud_offset(pgd, address); 2083 if (!pud_present(*pud)) 2084 goto out; 2085 2086 pmd = pmd_offset(pud, address); 2087 if (!pmd_present(*pmd) || pmd_trans_huge(*pmd)) 2088 goto out; 2089 2090 pte = pte_offset_map_lock(mm, pmd, address, &ptl); 2091 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR; 2092 _pte++, _address += PAGE_SIZE) { 2093 pte_t pteval = *_pte; 2094 if (pte_none(pteval)) { 2095 if (++none <= khugepaged_max_ptes_none) 2096 continue; 2097 else 2098 goto out_unmap; 2099 } 2100 if (!pte_present(pteval) || !pte_write(pteval)) 2101 goto out_unmap; 2102 page = vm_normal_page(vma, _address, pteval); 2103 if (unlikely(!page)) 2104 goto out_unmap; 2105 /* 2106 * Chose the node of the first page. This could 2107 * be more sophisticated and look at more pages, 2108 * but isn't for now. 2109 */ 2110 if (node == -1) 2111 node = page_to_nid(page); 2112 VM_BUG_ON(PageCompound(page)); 2113 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page)) 2114 goto out_unmap; 2115 /* cannot use mapcount: can't collapse if there's a gup pin */ 2116 if (page_count(page) != 1) 2117 goto out_unmap; 2118 if (pte_young(pteval) || PageReferenced(page) || 2119 mmu_notifier_test_young(vma->vm_mm, address)) 2120 referenced = 1; 2121 } 2122 if (referenced) 2123 ret = 1; 2124 out_unmap: 2125 pte_unmap_unlock(pte, ptl); 2126 if (ret) 2127 /* collapse_huge_page will return with the mmap_sem released */ 2128 collapse_huge_page(mm, address, hpage, vma, node); 2129 out: 2130 return ret; 2131 } 2132 2133 static void collect_mm_slot(struct mm_slot *mm_slot) 2134 { 2135 struct mm_struct *mm = mm_slot->mm; 2136 2137 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock)); 2138 2139 if (khugepaged_test_exit(mm)) { 2140 /* free mm_slot */ 2141 hlist_del(&mm_slot->hash); 2142 list_del(&mm_slot->mm_node); 2143 2144 /* 2145 * Not strictly needed because the mm exited already. 2146 * 2147 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags); 2148 */ 2149 2150 /* khugepaged_mm_lock actually not necessary for the below */ 2151 free_mm_slot(mm_slot); 2152 mmdrop(mm); 2153 } 2154 } 2155 2156 static unsigned int khugepaged_scan_mm_slot(unsigned int pages, 2157 struct page **hpage) 2158 __releases(&khugepaged_mm_lock) 2159 __acquires(&khugepaged_mm_lock) 2160 { 2161 struct mm_slot *mm_slot; 2162 struct mm_struct *mm; 2163 struct vm_area_struct *vma; 2164 int progress = 0; 2165 2166 VM_BUG_ON(!pages); 2167 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock)); 2168 2169 if (khugepaged_scan.mm_slot) 2170 mm_slot = khugepaged_scan.mm_slot; 2171 else { 2172 mm_slot = list_entry(khugepaged_scan.mm_head.next, 2173 struct mm_slot, mm_node); 2174 khugepaged_scan.address = 0; 2175 khugepaged_scan.mm_slot = mm_slot; 2176 } 2177 spin_unlock(&khugepaged_mm_lock); 2178 2179 mm = mm_slot->mm; 2180 down_read(&mm->mmap_sem); 2181 if (unlikely(khugepaged_test_exit(mm))) 2182 vma = NULL; 2183 else 2184 vma = find_vma(mm, khugepaged_scan.address); 2185 2186 progress++; 2187 for (; vma; vma = vma->vm_next) { 2188 unsigned long hstart, hend; 2189 2190 cond_resched(); 2191 if (unlikely(khugepaged_test_exit(mm))) { 2192 progress++; 2193 break; 2194 } 2195 2196 if ((!(vma->vm_flags & VM_HUGEPAGE) && 2197 !khugepaged_always()) || 2198 (vma->vm_flags & VM_NOHUGEPAGE)) { 2199 skip: 2200 progress++; 2201 continue; 2202 } 2203 if (!vma->anon_vma || vma->vm_ops) 2204 goto skip; 2205 if (is_vma_temporary_stack(vma)) 2206 goto skip; 2207 VM_BUG_ON(vma->vm_flags & VM_NO_THP); 2208 2209 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; 2210 hend = vma->vm_end & HPAGE_PMD_MASK; 2211 if (hstart >= hend) 2212 goto skip; 2213 if (khugepaged_scan.address > hend) 2214 goto skip; 2215 if (khugepaged_scan.address < hstart) 2216 khugepaged_scan.address = hstart; 2217 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK); 2218 2219 while (khugepaged_scan.address < hend) { 2220 int ret; 2221 cond_resched(); 2222 if (unlikely(khugepaged_test_exit(mm))) 2223 goto breakouterloop; 2224 2225 VM_BUG_ON(khugepaged_scan.address < hstart || 2226 khugepaged_scan.address + HPAGE_PMD_SIZE > 2227 hend); 2228 ret = khugepaged_scan_pmd(mm, vma, 2229 khugepaged_scan.address, 2230 hpage); 2231 /* move to next address */ 2232 khugepaged_scan.address += HPAGE_PMD_SIZE; 2233 progress += HPAGE_PMD_NR; 2234 if (ret) 2235 /* we released mmap_sem so break loop */ 2236 goto breakouterloop_mmap_sem; 2237 if (progress >= pages) 2238 goto breakouterloop; 2239 } 2240 } 2241 breakouterloop: 2242 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */ 2243 breakouterloop_mmap_sem: 2244 2245 spin_lock(&khugepaged_mm_lock); 2246 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot); 2247 /* 2248 * Release the current mm_slot if this mm is about to die, or 2249 * if we scanned all vmas of this mm. 2250 */ 2251 if (khugepaged_test_exit(mm) || !vma) { 2252 /* 2253 * Make sure that if mm_users is reaching zero while 2254 * khugepaged runs here, khugepaged_exit will find 2255 * mm_slot not pointing to the exiting mm. 2256 */ 2257 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) { 2258 khugepaged_scan.mm_slot = list_entry( 2259 mm_slot->mm_node.next, 2260 struct mm_slot, mm_node); 2261 khugepaged_scan.address = 0; 2262 } else { 2263 khugepaged_scan.mm_slot = NULL; 2264 khugepaged_full_scans++; 2265 } 2266 2267 collect_mm_slot(mm_slot); 2268 } 2269 2270 return progress; 2271 } 2272 2273 static int khugepaged_has_work(void) 2274 { 2275 return !list_empty(&khugepaged_scan.mm_head) && 2276 khugepaged_enabled(); 2277 } 2278 2279 static int khugepaged_wait_event(void) 2280 { 2281 return !list_empty(&khugepaged_scan.mm_head) || 2282 kthread_should_stop(); 2283 } 2284 2285 static void khugepaged_do_scan(void) 2286 { 2287 struct page *hpage = NULL; 2288 unsigned int progress = 0, pass_through_head = 0; 2289 unsigned int pages = khugepaged_pages_to_scan; 2290 bool wait = true; 2291 2292 barrier(); /* write khugepaged_pages_to_scan to local stack */ 2293 2294 while (progress < pages) { 2295 if (!khugepaged_prealloc_page(&hpage, &wait)) 2296 break; 2297 2298 cond_resched(); 2299 2300 if (unlikely(kthread_should_stop() || freezing(current))) 2301 break; 2302 2303 spin_lock(&khugepaged_mm_lock); 2304 if (!khugepaged_scan.mm_slot) 2305 pass_through_head++; 2306 if (khugepaged_has_work() && 2307 pass_through_head < 2) 2308 progress += khugepaged_scan_mm_slot(pages - progress, 2309 &hpage); 2310 else 2311 progress = pages; 2312 spin_unlock(&khugepaged_mm_lock); 2313 } 2314 2315 if (!IS_ERR_OR_NULL(hpage)) 2316 put_page(hpage); 2317 } 2318 2319 static void khugepaged_wait_work(void) 2320 { 2321 try_to_freeze(); 2322 2323 if (khugepaged_has_work()) { 2324 if (!khugepaged_scan_sleep_millisecs) 2325 return; 2326 2327 wait_event_freezable_timeout(khugepaged_wait, 2328 kthread_should_stop(), 2329 msecs_to_jiffies(khugepaged_scan_sleep_millisecs)); 2330 return; 2331 } 2332 2333 if (khugepaged_enabled()) 2334 wait_event_freezable(khugepaged_wait, khugepaged_wait_event()); 2335 } 2336 2337 static int khugepaged(void *none) 2338 { 2339 struct mm_slot *mm_slot; 2340 2341 set_freezable(); 2342 set_user_nice(current, 19); 2343 2344 while (!kthread_should_stop()) { 2345 khugepaged_do_scan(); 2346 khugepaged_wait_work(); 2347 } 2348 2349 spin_lock(&khugepaged_mm_lock); 2350 mm_slot = khugepaged_scan.mm_slot; 2351 khugepaged_scan.mm_slot = NULL; 2352 if (mm_slot) 2353 collect_mm_slot(mm_slot); 2354 spin_unlock(&khugepaged_mm_lock); 2355 return 0; 2356 } 2357 2358 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd) 2359 { 2360 struct page *page; 2361 2362 spin_lock(&mm->page_table_lock); 2363 if (unlikely(!pmd_trans_huge(*pmd))) { 2364 spin_unlock(&mm->page_table_lock); 2365 return; 2366 } 2367 page = pmd_page(*pmd); 2368 VM_BUG_ON(!page_count(page)); 2369 get_page(page); 2370 spin_unlock(&mm->page_table_lock); 2371 2372 split_huge_page(page); 2373 2374 put_page(page); 2375 BUG_ON(pmd_trans_huge(*pmd)); 2376 } 2377 2378 static void split_huge_page_address(struct mm_struct *mm, 2379 unsigned long address) 2380 { 2381 pgd_t *pgd; 2382 pud_t *pud; 2383 pmd_t *pmd; 2384 2385 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK)); 2386 2387 pgd = pgd_offset(mm, address); 2388 if (!pgd_present(*pgd)) 2389 return; 2390 2391 pud = pud_offset(pgd, address); 2392 if (!pud_present(*pud)) 2393 return; 2394 2395 pmd = pmd_offset(pud, address); 2396 if (!pmd_present(*pmd)) 2397 return; 2398 /* 2399 * Caller holds the mmap_sem write mode, so a huge pmd cannot 2400 * materialize from under us. 2401 */ 2402 split_huge_page_pmd(mm, pmd); 2403 } 2404 2405 void __vma_adjust_trans_huge(struct vm_area_struct *vma, 2406 unsigned long start, 2407 unsigned long end, 2408 long adjust_next) 2409 { 2410 /* 2411 * If the new start address isn't hpage aligned and it could 2412 * previously contain an hugepage: check if we need to split 2413 * an huge pmd. 2414 */ 2415 if (start & ~HPAGE_PMD_MASK && 2416 (start & HPAGE_PMD_MASK) >= vma->vm_start && 2417 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) 2418 split_huge_page_address(vma->vm_mm, start); 2419 2420 /* 2421 * If the new end address isn't hpage aligned and it could 2422 * previously contain an hugepage: check if we need to split 2423 * an huge pmd. 2424 */ 2425 if (end & ~HPAGE_PMD_MASK && 2426 (end & HPAGE_PMD_MASK) >= vma->vm_start && 2427 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) 2428 split_huge_page_address(vma->vm_mm, end); 2429 2430 /* 2431 * If we're also updating the vma->vm_next->vm_start, if the new 2432 * vm_next->vm_start isn't page aligned and it could previously 2433 * contain an hugepage: check if we need to split an huge pmd. 2434 */ 2435 if (adjust_next > 0) { 2436 struct vm_area_struct *next = vma->vm_next; 2437 unsigned long nstart = next->vm_start; 2438 nstart += adjust_next << PAGE_SHIFT; 2439 if (nstart & ~HPAGE_PMD_MASK && 2440 (nstart & HPAGE_PMD_MASK) >= next->vm_start && 2441 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end) 2442 split_huge_page_address(next->vm_mm, nstart); 2443 } 2444 } 2445