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/shrinker.h> 16 #include <linux/mm_inline.h> 17 #include <linux/kthread.h> 18 #include <linux/khugepaged.h> 19 #include <linux/freezer.h> 20 #include <linux/mman.h> 21 #include <linux/pagemap.h> 22 #include <linux/migrate.h> 23 #include <linux/hashtable.h> 24 25 #include <asm/tlb.h> 26 #include <asm/pgalloc.h> 27 #include "internal.h" 28 29 /* 30 * By default transparent hugepage support is disabled in order that avoid 31 * to risk increase the memory footprint of applications without a guaranteed 32 * benefit. When transparent hugepage support is enabled, is for all mappings, 33 * and khugepaged scans all mappings. 34 * Defrag is invoked by khugepaged hugepage allocations and by page faults 35 * for all hugepage allocations. 36 */ 37 unsigned long transparent_hugepage_flags __read_mostly = 38 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS 39 (1<<TRANSPARENT_HUGEPAGE_FLAG)| 40 #endif 41 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE 42 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)| 43 #endif 44 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)| 45 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)| 46 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 47 48 /* default scan 8*512 pte (or vmas) every 30 second */ 49 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8; 50 static unsigned int khugepaged_pages_collapsed; 51 static unsigned int khugepaged_full_scans; 52 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000; 53 /* during fragmentation poll the hugepage allocator once every minute */ 54 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000; 55 static struct task_struct *khugepaged_thread __read_mostly; 56 static DEFINE_MUTEX(khugepaged_mutex); 57 static DEFINE_SPINLOCK(khugepaged_mm_lock); 58 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait); 59 /* 60 * default collapse hugepages if there is at least one pte mapped like 61 * it would have happened if the vma was large enough during page 62 * fault. 63 */ 64 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1; 65 66 static int khugepaged(void *none); 67 static int khugepaged_slab_init(void); 68 69 #define MM_SLOTS_HASH_BITS 10 70 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); 71 72 static struct kmem_cache *mm_slot_cache __read_mostly; 73 74 /** 75 * struct mm_slot - hash lookup from mm to mm_slot 76 * @hash: hash collision list 77 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head 78 * @mm: the mm that this information is valid for 79 */ 80 struct mm_slot { 81 struct hlist_node hash; 82 struct list_head mm_node; 83 struct mm_struct *mm; 84 }; 85 86 /** 87 * struct khugepaged_scan - cursor for scanning 88 * @mm_head: the head of the mm list to scan 89 * @mm_slot: the current mm_slot we are scanning 90 * @address: the next address inside that to be scanned 91 * 92 * There is only the one khugepaged_scan instance of this cursor structure. 93 */ 94 struct khugepaged_scan { 95 struct list_head mm_head; 96 struct mm_slot *mm_slot; 97 unsigned long address; 98 }; 99 static struct khugepaged_scan khugepaged_scan = { 100 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head), 101 }; 102 103 104 static int set_recommended_min_free_kbytes(void) 105 { 106 struct zone *zone; 107 int nr_zones = 0; 108 unsigned long recommended_min; 109 110 if (!khugepaged_enabled()) 111 return 0; 112 113 for_each_populated_zone(zone) 114 nr_zones++; 115 116 /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */ 117 recommended_min = pageblock_nr_pages * nr_zones * 2; 118 119 /* 120 * Make sure that on average at least two pageblocks are almost free 121 * of another type, one for a migratetype to fall back to and a 122 * second to avoid subsequent fallbacks of other types There are 3 123 * MIGRATE_TYPES we care about. 124 */ 125 recommended_min += pageblock_nr_pages * nr_zones * 126 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES; 127 128 /* don't ever allow to reserve more than 5% of the lowmem */ 129 recommended_min = min(recommended_min, 130 (unsigned long) nr_free_buffer_pages() / 20); 131 recommended_min <<= (PAGE_SHIFT-10); 132 133 if (recommended_min > min_free_kbytes) 134 min_free_kbytes = recommended_min; 135 setup_per_zone_wmarks(); 136 return 0; 137 } 138 late_initcall(set_recommended_min_free_kbytes); 139 140 static int start_khugepaged(void) 141 { 142 int err = 0; 143 if (khugepaged_enabled()) { 144 if (!khugepaged_thread) 145 khugepaged_thread = kthread_run(khugepaged, NULL, 146 "khugepaged"); 147 if (unlikely(IS_ERR(khugepaged_thread))) { 148 printk(KERN_ERR 149 "khugepaged: kthread_run(khugepaged) failed\n"); 150 err = PTR_ERR(khugepaged_thread); 151 khugepaged_thread = NULL; 152 } 153 154 if (!list_empty(&khugepaged_scan.mm_head)) 155 wake_up_interruptible(&khugepaged_wait); 156 157 set_recommended_min_free_kbytes(); 158 } else if (khugepaged_thread) { 159 kthread_stop(khugepaged_thread); 160 khugepaged_thread = NULL; 161 } 162 163 return err; 164 } 165 166 static atomic_t huge_zero_refcount; 167 static struct page *huge_zero_page __read_mostly; 168 169 static inline bool is_huge_zero_page(struct page *page) 170 { 171 return ACCESS_ONCE(huge_zero_page) == page; 172 } 173 174 static inline bool is_huge_zero_pmd(pmd_t pmd) 175 { 176 return is_huge_zero_page(pmd_page(pmd)); 177 } 178 179 static struct page *get_huge_zero_page(void) 180 { 181 struct page *zero_page; 182 retry: 183 if (likely(atomic_inc_not_zero(&huge_zero_refcount))) 184 return ACCESS_ONCE(huge_zero_page); 185 186 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE, 187 HPAGE_PMD_ORDER); 188 if (!zero_page) { 189 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED); 190 return NULL; 191 } 192 count_vm_event(THP_ZERO_PAGE_ALLOC); 193 preempt_disable(); 194 if (cmpxchg(&huge_zero_page, NULL, zero_page)) { 195 preempt_enable(); 196 __free_page(zero_page); 197 goto retry; 198 } 199 200 /* We take additional reference here. It will be put back by shrinker */ 201 atomic_set(&huge_zero_refcount, 2); 202 preempt_enable(); 203 return ACCESS_ONCE(huge_zero_page); 204 } 205 206 static void put_huge_zero_page(void) 207 { 208 /* 209 * Counter should never go to zero here. Only shrinker can put 210 * last reference. 211 */ 212 BUG_ON(atomic_dec_and_test(&huge_zero_refcount)); 213 } 214 215 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink, 216 struct shrink_control *sc) 217 { 218 /* we can free zero page only if last reference remains */ 219 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0; 220 } 221 222 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink, 223 struct shrink_control *sc) 224 { 225 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) { 226 struct page *zero_page = xchg(&huge_zero_page, NULL); 227 BUG_ON(zero_page == NULL); 228 __free_page(zero_page); 229 return HPAGE_PMD_NR; 230 } 231 232 return 0; 233 } 234 235 static struct shrinker huge_zero_page_shrinker = { 236 .count_objects = shrink_huge_zero_page_count, 237 .scan_objects = shrink_huge_zero_page_scan, 238 .seeks = DEFAULT_SEEKS, 239 }; 240 241 #ifdef CONFIG_SYSFS 242 243 static ssize_t double_flag_show(struct kobject *kobj, 244 struct kobj_attribute *attr, char *buf, 245 enum transparent_hugepage_flag enabled, 246 enum transparent_hugepage_flag req_madv) 247 { 248 if (test_bit(enabled, &transparent_hugepage_flags)) { 249 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags)); 250 return sprintf(buf, "[always] madvise never\n"); 251 } else if (test_bit(req_madv, &transparent_hugepage_flags)) 252 return sprintf(buf, "always [madvise] never\n"); 253 else 254 return sprintf(buf, "always madvise [never]\n"); 255 } 256 static ssize_t double_flag_store(struct kobject *kobj, 257 struct kobj_attribute *attr, 258 const char *buf, size_t count, 259 enum transparent_hugepage_flag enabled, 260 enum transparent_hugepage_flag req_madv) 261 { 262 if (!memcmp("always", buf, 263 min(sizeof("always")-1, count))) { 264 set_bit(enabled, &transparent_hugepage_flags); 265 clear_bit(req_madv, &transparent_hugepage_flags); 266 } else if (!memcmp("madvise", buf, 267 min(sizeof("madvise")-1, count))) { 268 clear_bit(enabled, &transparent_hugepage_flags); 269 set_bit(req_madv, &transparent_hugepage_flags); 270 } else if (!memcmp("never", buf, 271 min(sizeof("never")-1, count))) { 272 clear_bit(enabled, &transparent_hugepage_flags); 273 clear_bit(req_madv, &transparent_hugepage_flags); 274 } else 275 return -EINVAL; 276 277 return count; 278 } 279 280 static ssize_t enabled_show(struct kobject *kobj, 281 struct kobj_attribute *attr, char *buf) 282 { 283 return double_flag_show(kobj, attr, buf, 284 TRANSPARENT_HUGEPAGE_FLAG, 285 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG); 286 } 287 static ssize_t enabled_store(struct kobject *kobj, 288 struct kobj_attribute *attr, 289 const char *buf, size_t count) 290 { 291 ssize_t ret; 292 293 ret = double_flag_store(kobj, attr, buf, count, 294 TRANSPARENT_HUGEPAGE_FLAG, 295 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG); 296 297 if (ret > 0) { 298 int err; 299 300 mutex_lock(&khugepaged_mutex); 301 err = start_khugepaged(); 302 mutex_unlock(&khugepaged_mutex); 303 304 if (err) 305 ret = err; 306 } 307 308 return ret; 309 } 310 static struct kobj_attribute enabled_attr = 311 __ATTR(enabled, 0644, enabled_show, enabled_store); 312 313 static ssize_t single_flag_show(struct kobject *kobj, 314 struct kobj_attribute *attr, char *buf, 315 enum transparent_hugepage_flag flag) 316 { 317 return sprintf(buf, "%d\n", 318 !!test_bit(flag, &transparent_hugepage_flags)); 319 } 320 321 static ssize_t single_flag_store(struct kobject *kobj, 322 struct kobj_attribute *attr, 323 const char *buf, size_t count, 324 enum transparent_hugepage_flag flag) 325 { 326 unsigned long value; 327 int ret; 328 329 ret = kstrtoul(buf, 10, &value); 330 if (ret < 0) 331 return ret; 332 if (value > 1) 333 return -EINVAL; 334 335 if (value) 336 set_bit(flag, &transparent_hugepage_flags); 337 else 338 clear_bit(flag, &transparent_hugepage_flags); 339 340 return count; 341 } 342 343 /* 344 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind 345 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of 346 * memory just to allocate one more hugepage. 347 */ 348 static ssize_t defrag_show(struct kobject *kobj, 349 struct kobj_attribute *attr, char *buf) 350 { 351 return double_flag_show(kobj, attr, buf, 352 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG, 353 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG); 354 } 355 static ssize_t defrag_store(struct kobject *kobj, 356 struct kobj_attribute *attr, 357 const char *buf, size_t count) 358 { 359 return double_flag_store(kobj, attr, buf, count, 360 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG, 361 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG); 362 } 363 static struct kobj_attribute defrag_attr = 364 __ATTR(defrag, 0644, defrag_show, defrag_store); 365 366 static ssize_t use_zero_page_show(struct kobject *kobj, 367 struct kobj_attribute *attr, char *buf) 368 { 369 return single_flag_show(kobj, attr, buf, 370 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 371 } 372 static ssize_t use_zero_page_store(struct kobject *kobj, 373 struct kobj_attribute *attr, const char *buf, size_t count) 374 { 375 return single_flag_store(kobj, attr, buf, count, 376 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 377 } 378 static struct kobj_attribute use_zero_page_attr = 379 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store); 380 #ifdef CONFIG_DEBUG_VM 381 static ssize_t debug_cow_show(struct kobject *kobj, 382 struct kobj_attribute *attr, char *buf) 383 { 384 return single_flag_show(kobj, attr, buf, 385 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); 386 } 387 static ssize_t debug_cow_store(struct kobject *kobj, 388 struct kobj_attribute *attr, 389 const char *buf, size_t count) 390 { 391 return single_flag_store(kobj, attr, buf, count, 392 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); 393 } 394 static struct kobj_attribute debug_cow_attr = 395 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store); 396 #endif /* CONFIG_DEBUG_VM */ 397 398 static struct attribute *hugepage_attr[] = { 399 &enabled_attr.attr, 400 &defrag_attr.attr, 401 &use_zero_page_attr.attr, 402 #ifdef CONFIG_DEBUG_VM 403 &debug_cow_attr.attr, 404 #endif 405 NULL, 406 }; 407 408 static struct attribute_group hugepage_attr_group = { 409 .attrs = hugepage_attr, 410 }; 411 412 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj, 413 struct kobj_attribute *attr, 414 char *buf) 415 { 416 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs); 417 } 418 419 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj, 420 struct kobj_attribute *attr, 421 const char *buf, size_t count) 422 { 423 unsigned long msecs; 424 int err; 425 426 err = kstrtoul(buf, 10, &msecs); 427 if (err || msecs > UINT_MAX) 428 return -EINVAL; 429 430 khugepaged_scan_sleep_millisecs = msecs; 431 wake_up_interruptible(&khugepaged_wait); 432 433 return count; 434 } 435 static struct kobj_attribute scan_sleep_millisecs_attr = 436 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show, 437 scan_sleep_millisecs_store); 438 439 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj, 440 struct kobj_attribute *attr, 441 char *buf) 442 { 443 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs); 444 } 445 446 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj, 447 struct kobj_attribute *attr, 448 const char *buf, size_t count) 449 { 450 unsigned long msecs; 451 int err; 452 453 err = kstrtoul(buf, 10, &msecs); 454 if (err || msecs > UINT_MAX) 455 return -EINVAL; 456 457 khugepaged_alloc_sleep_millisecs = msecs; 458 wake_up_interruptible(&khugepaged_wait); 459 460 return count; 461 } 462 static struct kobj_attribute alloc_sleep_millisecs_attr = 463 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show, 464 alloc_sleep_millisecs_store); 465 466 static ssize_t pages_to_scan_show(struct kobject *kobj, 467 struct kobj_attribute *attr, 468 char *buf) 469 { 470 return sprintf(buf, "%u\n", khugepaged_pages_to_scan); 471 } 472 static ssize_t pages_to_scan_store(struct kobject *kobj, 473 struct kobj_attribute *attr, 474 const char *buf, size_t count) 475 { 476 int err; 477 unsigned long pages; 478 479 err = kstrtoul(buf, 10, &pages); 480 if (err || !pages || pages > UINT_MAX) 481 return -EINVAL; 482 483 khugepaged_pages_to_scan = pages; 484 485 return count; 486 } 487 static struct kobj_attribute pages_to_scan_attr = 488 __ATTR(pages_to_scan, 0644, pages_to_scan_show, 489 pages_to_scan_store); 490 491 static ssize_t pages_collapsed_show(struct kobject *kobj, 492 struct kobj_attribute *attr, 493 char *buf) 494 { 495 return sprintf(buf, "%u\n", khugepaged_pages_collapsed); 496 } 497 static struct kobj_attribute pages_collapsed_attr = 498 __ATTR_RO(pages_collapsed); 499 500 static ssize_t full_scans_show(struct kobject *kobj, 501 struct kobj_attribute *attr, 502 char *buf) 503 { 504 return sprintf(buf, "%u\n", khugepaged_full_scans); 505 } 506 static struct kobj_attribute full_scans_attr = 507 __ATTR_RO(full_scans); 508 509 static ssize_t khugepaged_defrag_show(struct kobject *kobj, 510 struct kobj_attribute *attr, char *buf) 511 { 512 return single_flag_show(kobj, attr, buf, 513 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); 514 } 515 static ssize_t khugepaged_defrag_store(struct kobject *kobj, 516 struct kobj_attribute *attr, 517 const char *buf, size_t count) 518 { 519 return single_flag_store(kobj, attr, buf, count, 520 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); 521 } 522 static struct kobj_attribute khugepaged_defrag_attr = 523 __ATTR(defrag, 0644, khugepaged_defrag_show, 524 khugepaged_defrag_store); 525 526 /* 527 * max_ptes_none controls if khugepaged should collapse hugepages over 528 * any unmapped ptes in turn potentially increasing the memory 529 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not 530 * reduce the available free memory in the system as it 531 * runs. Increasing max_ptes_none will instead potentially reduce the 532 * free memory in the system during the khugepaged scan. 533 */ 534 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj, 535 struct kobj_attribute *attr, 536 char *buf) 537 { 538 return sprintf(buf, "%u\n", khugepaged_max_ptes_none); 539 } 540 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj, 541 struct kobj_attribute *attr, 542 const char *buf, size_t count) 543 { 544 int err; 545 unsigned long max_ptes_none; 546 547 err = kstrtoul(buf, 10, &max_ptes_none); 548 if (err || max_ptes_none > HPAGE_PMD_NR-1) 549 return -EINVAL; 550 551 khugepaged_max_ptes_none = max_ptes_none; 552 553 return count; 554 } 555 static struct kobj_attribute khugepaged_max_ptes_none_attr = 556 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show, 557 khugepaged_max_ptes_none_store); 558 559 static struct attribute *khugepaged_attr[] = { 560 &khugepaged_defrag_attr.attr, 561 &khugepaged_max_ptes_none_attr.attr, 562 &pages_to_scan_attr.attr, 563 &pages_collapsed_attr.attr, 564 &full_scans_attr.attr, 565 &scan_sleep_millisecs_attr.attr, 566 &alloc_sleep_millisecs_attr.attr, 567 NULL, 568 }; 569 570 static struct attribute_group khugepaged_attr_group = { 571 .attrs = khugepaged_attr, 572 .name = "khugepaged", 573 }; 574 575 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj) 576 { 577 int err; 578 579 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj); 580 if (unlikely(!*hugepage_kobj)) { 581 printk(KERN_ERR "hugepage: failed to create transparent hugepage kobject\n"); 582 return -ENOMEM; 583 } 584 585 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group); 586 if (err) { 587 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n"); 588 goto delete_obj; 589 } 590 591 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group); 592 if (err) { 593 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n"); 594 goto remove_hp_group; 595 } 596 597 return 0; 598 599 remove_hp_group: 600 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group); 601 delete_obj: 602 kobject_put(*hugepage_kobj); 603 return err; 604 } 605 606 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj) 607 { 608 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group); 609 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group); 610 kobject_put(hugepage_kobj); 611 } 612 #else 613 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj) 614 { 615 return 0; 616 } 617 618 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj) 619 { 620 } 621 #endif /* CONFIG_SYSFS */ 622 623 static int __init hugepage_init(void) 624 { 625 int err; 626 struct kobject *hugepage_kobj; 627 628 if (!has_transparent_hugepage()) { 629 transparent_hugepage_flags = 0; 630 return -EINVAL; 631 } 632 633 err = hugepage_init_sysfs(&hugepage_kobj); 634 if (err) 635 return err; 636 637 err = khugepaged_slab_init(); 638 if (err) 639 goto out; 640 641 register_shrinker(&huge_zero_page_shrinker); 642 643 /* 644 * By default disable transparent hugepages on smaller systems, 645 * where the extra memory used could hurt more than TLB overhead 646 * is likely to save. The admin can still enable it through /sys. 647 */ 648 if (totalram_pages < (512 << (20 - PAGE_SHIFT))) 649 transparent_hugepage_flags = 0; 650 651 start_khugepaged(); 652 653 return 0; 654 out: 655 hugepage_exit_sysfs(hugepage_kobj); 656 return err; 657 } 658 module_init(hugepage_init) 659 660 static int __init setup_transparent_hugepage(char *str) 661 { 662 int ret = 0; 663 if (!str) 664 goto out; 665 if (!strcmp(str, "always")) { 666 set_bit(TRANSPARENT_HUGEPAGE_FLAG, 667 &transparent_hugepage_flags); 668 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 669 &transparent_hugepage_flags); 670 ret = 1; 671 } else if (!strcmp(str, "madvise")) { 672 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 673 &transparent_hugepage_flags); 674 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 675 &transparent_hugepage_flags); 676 ret = 1; 677 } else if (!strcmp(str, "never")) { 678 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 679 &transparent_hugepage_flags); 680 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 681 &transparent_hugepage_flags); 682 ret = 1; 683 } 684 out: 685 if (!ret) 686 printk(KERN_WARNING 687 "transparent_hugepage= cannot parse, ignored\n"); 688 return ret; 689 } 690 __setup("transparent_hugepage=", setup_transparent_hugepage); 691 692 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma) 693 { 694 if (likely(vma->vm_flags & VM_WRITE)) 695 pmd = pmd_mkwrite(pmd); 696 return pmd; 697 } 698 699 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot) 700 { 701 pmd_t entry; 702 entry = mk_pmd(page, prot); 703 entry = pmd_mkhuge(entry); 704 return entry; 705 } 706 707 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm, 708 struct vm_area_struct *vma, 709 unsigned long haddr, pmd_t *pmd, 710 struct page *page) 711 { 712 pgtable_t pgtable; 713 spinlock_t *ptl; 714 715 VM_BUG_ON(!PageCompound(page)); 716 pgtable = pte_alloc_one(mm, haddr); 717 if (unlikely(!pgtable)) 718 return VM_FAULT_OOM; 719 720 clear_huge_page(page, haddr, HPAGE_PMD_NR); 721 /* 722 * The memory barrier inside __SetPageUptodate makes sure that 723 * clear_huge_page writes become visible before the set_pmd_at() 724 * write. 725 */ 726 __SetPageUptodate(page); 727 728 ptl = pmd_lock(mm, pmd); 729 if (unlikely(!pmd_none(*pmd))) { 730 spin_unlock(ptl); 731 mem_cgroup_uncharge_page(page); 732 put_page(page); 733 pte_free(mm, pgtable); 734 } else { 735 pmd_t entry; 736 entry = mk_huge_pmd(page, vma->vm_page_prot); 737 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 738 page_add_new_anon_rmap(page, vma, haddr); 739 pgtable_trans_huge_deposit(mm, pmd, pgtable); 740 set_pmd_at(mm, haddr, pmd, entry); 741 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR); 742 atomic_long_inc(&mm->nr_ptes); 743 spin_unlock(ptl); 744 } 745 746 return 0; 747 } 748 749 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp) 750 { 751 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp; 752 } 753 754 static inline struct page *alloc_hugepage_vma(int defrag, 755 struct vm_area_struct *vma, 756 unsigned long haddr, int nd, 757 gfp_t extra_gfp) 758 { 759 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp), 760 HPAGE_PMD_ORDER, vma, haddr, nd); 761 } 762 763 /* Caller must hold page table lock. */ 764 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, 765 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, 766 struct page *zero_page) 767 { 768 pmd_t entry; 769 if (!pmd_none(*pmd)) 770 return false; 771 entry = mk_pmd(zero_page, vma->vm_page_prot); 772 entry = pmd_wrprotect(entry); 773 entry = pmd_mkhuge(entry); 774 pgtable_trans_huge_deposit(mm, pmd, pgtable); 775 set_pmd_at(mm, haddr, pmd, entry); 776 atomic_long_inc(&mm->nr_ptes); 777 return true; 778 } 779 780 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma, 781 unsigned long address, pmd_t *pmd, 782 unsigned int flags) 783 { 784 struct page *page; 785 unsigned long haddr = address & HPAGE_PMD_MASK; 786 787 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end) 788 return VM_FAULT_FALLBACK; 789 if (unlikely(anon_vma_prepare(vma))) 790 return VM_FAULT_OOM; 791 if (unlikely(khugepaged_enter(vma))) 792 return VM_FAULT_OOM; 793 if (!(flags & FAULT_FLAG_WRITE) && 794 transparent_hugepage_use_zero_page()) { 795 spinlock_t *ptl; 796 pgtable_t pgtable; 797 struct page *zero_page; 798 bool set; 799 pgtable = pte_alloc_one(mm, haddr); 800 if (unlikely(!pgtable)) 801 return VM_FAULT_OOM; 802 zero_page = get_huge_zero_page(); 803 if (unlikely(!zero_page)) { 804 pte_free(mm, pgtable); 805 count_vm_event(THP_FAULT_FALLBACK); 806 return VM_FAULT_FALLBACK; 807 } 808 ptl = pmd_lock(mm, pmd); 809 set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd, 810 zero_page); 811 spin_unlock(ptl); 812 if (!set) { 813 pte_free(mm, pgtable); 814 put_huge_zero_page(); 815 } 816 return 0; 817 } 818 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma), 819 vma, haddr, numa_node_id(), 0); 820 if (unlikely(!page)) { 821 count_vm_event(THP_FAULT_FALLBACK); 822 return VM_FAULT_FALLBACK; 823 } 824 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) { 825 put_page(page); 826 count_vm_event(THP_FAULT_FALLBACK); 827 return VM_FAULT_FALLBACK; 828 } 829 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page))) { 830 mem_cgroup_uncharge_page(page); 831 put_page(page); 832 count_vm_event(THP_FAULT_FALLBACK); 833 return VM_FAULT_FALLBACK; 834 } 835 836 count_vm_event(THP_FAULT_ALLOC); 837 return 0; 838 } 839 840 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, 841 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, 842 struct vm_area_struct *vma) 843 { 844 spinlock_t *dst_ptl, *src_ptl; 845 struct page *src_page; 846 pmd_t pmd; 847 pgtable_t pgtable; 848 int ret; 849 850 ret = -ENOMEM; 851 pgtable = pte_alloc_one(dst_mm, addr); 852 if (unlikely(!pgtable)) 853 goto out; 854 855 dst_ptl = pmd_lock(dst_mm, dst_pmd); 856 src_ptl = pmd_lockptr(src_mm, src_pmd); 857 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 858 859 ret = -EAGAIN; 860 pmd = *src_pmd; 861 if (unlikely(!pmd_trans_huge(pmd))) { 862 pte_free(dst_mm, pgtable); 863 goto out_unlock; 864 } 865 /* 866 * When page table lock is held, the huge zero pmd should not be 867 * under splitting since we don't split the page itself, only pmd to 868 * a page table. 869 */ 870 if (is_huge_zero_pmd(pmd)) { 871 struct page *zero_page; 872 bool set; 873 /* 874 * get_huge_zero_page() will never allocate a new page here, 875 * since we already have a zero page to copy. It just takes a 876 * reference. 877 */ 878 zero_page = get_huge_zero_page(); 879 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd, 880 zero_page); 881 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */ 882 ret = 0; 883 goto out_unlock; 884 } 885 886 /* mmap_sem prevents this happening but warn if that changes */ 887 WARN_ON(pmd_trans_migrating(pmd)); 888 889 if (unlikely(pmd_trans_splitting(pmd))) { 890 /* split huge page running from under us */ 891 spin_unlock(src_ptl); 892 spin_unlock(dst_ptl); 893 pte_free(dst_mm, pgtable); 894 895 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */ 896 goto out; 897 } 898 src_page = pmd_page(pmd); 899 VM_BUG_ON(!PageHead(src_page)); 900 get_page(src_page); 901 page_dup_rmap(src_page); 902 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 903 904 pmdp_set_wrprotect(src_mm, addr, src_pmd); 905 pmd = pmd_mkold(pmd_wrprotect(pmd)); 906 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); 907 set_pmd_at(dst_mm, addr, dst_pmd, pmd); 908 atomic_long_inc(&dst_mm->nr_ptes); 909 910 ret = 0; 911 out_unlock: 912 spin_unlock(src_ptl); 913 spin_unlock(dst_ptl); 914 out: 915 return ret; 916 } 917 918 void huge_pmd_set_accessed(struct mm_struct *mm, 919 struct vm_area_struct *vma, 920 unsigned long address, 921 pmd_t *pmd, pmd_t orig_pmd, 922 int dirty) 923 { 924 spinlock_t *ptl; 925 pmd_t entry; 926 unsigned long haddr; 927 928 ptl = pmd_lock(mm, pmd); 929 if (unlikely(!pmd_same(*pmd, orig_pmd))) 930 goto unlock; 931 932 entry = pmd_mkyoung(orig_pmd); 933 haddr = address & HPAGE_PMD_MASK; 934 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty)) 935 update_mmu_cache_pmd(vma, address, pmd); 936 937 unlock: 938 spin_unlock(ptl); 939 } 940 941 static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct *mm, 942 struct vm_area_struct *vma, unsigned long address, 943 pmd_t *pmd, pmd_t orig_pmd, unsigned long haddr) 944 { 945 spinlock_t *ptl; 946 pgtable_t pgtable; 947 pmd_t _pmd; 948 struct page *page; 949 int i, ret = 0; 950 unsigned long mmun_start; /* For mmu_notifiers */ 951 unsigned long mmun_end; /* For mmu_notifiers */ 952 953 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); 954 if (!page) { 955 ret |= VM_FAULT_OOM; 956 goto out; 957 } 958 959 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) { 960 put_page(page); 961 ret |= VM_FAULT_OOM; 962 goto out; 963 } 964 965 clear_user_highpage(page, address); 966 __SetPageUptodate(page); 967 968 mmun_start = haddr; 969 mmun_end = haddr + HPAGE_PMD_SIZE; 970 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 971 972 ptl = pmd_lock(mm, pmd); 973 if (unlikely(!pmd_same(*pmd, orig_pmd))) 974 goto out_free_page; 975 976 pmdp_clear_flush(vma, haddr, pmd); 977 /* leave pmd empty until pte is filled */ 978 979 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 980 pmd_populate(mm, &_pmd, pgtable); 981 982 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 983 pte_t *pte, entry; 984 if (haddr == (address & PAGE_MASK)) { 985 entry = mk_pte(page, vma->vm_page_prot); 986 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 987 page_add_new_anon_rmap(page, vma, haddr); 988 } else { 989 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot); 990 entry = pte_mkspecial(entry); 991 } 992 pte = pte_offset_map(&_pmd, haddr); 993 VM_BUG_ON(!pte_none(*pte)); 994 set_pte_at(mm, haddr, pte, entry); 995 pte_unmap(pte); 996 } 997 smp_wmb(); /* make pte visible before pmd */ 998 pmd_populate(mm, pmd, pgtable); 999 spin_unlock(ptl); 1000 put_huge_zero_page(); 1001 inc_mm_counter(mm, MM_ANONPAGES); 1002 1003 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 1004 1005 ret |= VM_FAULT_WRITE; 1006 out: 1007 return ret; 1008 out_free_page: 1009 spin_unlock(ptl); 1010 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 1011 mem_cgroup_uncharge_page(page); 1012 put_page(page); 1013 goto out; 1014 } 1015 1016 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm, 1017 struct vm_area_struct *vma, 1018 unsigned long address, 1019 pmd_t *pmd, pmd_t orig_pmd, 1020 struct page *page, 1021 unsigned long haddr) 1022 { 1023 spinlock_t *ptl; 1024 pgtable_t pgtable; 1025 pmd_t _pmd; 1026 int ret = 0, i; 1027 struct page **pages; 1028 unsigned long mmun_start; /* For mmu_notifiers */ 1029 unsigned long mmun_end; /* For mmu_notifiers */ 1030 1031 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR, 1032 GFP_KERNEL); 1033 if (unlikely(!pages)) { 1034 ret |= VM_FAULT_OOM; 1035 goto out; 1036 } 1037 1038 for (i = 0; i < HPAGE_PMD_NR; i++) { 1039 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE | 1040 __GFP_OTHER_NODE, 1041 vma, address, page_to_nid(page)); 1042 if (unlikely(!pages[i] || 1043 mem_cgroup_newpage_charge(pages[i], mm, 1044 GFP_KERNEL))) { 1045 if (pages[i]) 1046 put_page(pages[i]); 1047 mem_cgroup_uncharge_start(); 1048 while (--i >= 0) { 1049 mem_cgroup_uncharge_page(pages[i]); 1050 put_page(pages[i]); 1051 } 1052 mem_cgroup_uncharge_end(); 1053 kfree(pages); 1054 ret |= VM_FAULT_OOM; 1055 goto out; 1056 } 1057 } 1058 1059 for (i = 0; i < HPAGE_PMD_NR; i++) { 1060 copy_user_highpage(pages[i], page + i, 1061 haddr + PAGE_SIZE * i, vma); 1062 __SetPageUptodate(pages[i]); 1063 cond_resched(); 1064 } 1065 1066 mmun_start = haddr; 1067 mmun_end = haddr + HPAGE_PMD_SIZE; 1068 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 1069 1070 ptl = pmd_lock(mm, pmd); 1071 if (unlikely(!pmd_same(*pmd, orig_pmd))) 1072 goto out_free_pages; 1073 VM_BUG_ON(!PageHead(page)); 1074 1075 pmdp_clear_flush(vma, haddr, pmd); 1076 /* leave pmd empty until pte is filled */ 1077 1078 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 1079 pmd_populate(mm, &_pmd, pgtable); 1080 1081 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 1082 pte_t *pte, entry; 1083 entry = mk_pte(pages[i], vma->vm_page_prot); 1084 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 1085 page_add_new_anon_rmap(pages[i], vma, haddr); 1086 pte = pte_offset_map(&_pmd, haddr); 1087 VM_BUG_ON(!pte_none(*pte)); 1088 set_pte_at(mm, haddr, pte, entry); 1089 pte_unmap(pte); 1090 } 1091 kfree(pages); 1092 1093 smp_wmb(); /* make pte visible before pmd */ 1094 pmd_populate(mm, pmd, pgtable); 1095 page_remove_rmap(page); 1096 spin_unlock(ptl); 1097 1098 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 1099 1100 ret |= VM_FAULT_WRITE; 1101 put_page(page); 1102 1103 out: 1104 return ret; 1105 1106 out_free_pages: 1107 spin_unlock(ptl); 1108 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 1109 mem_cgroup_uncharge_start(); 1110 for (i = 0; i < HPAGE_PMD_NR; i++) { 1111 mem_cgroup_uncharge_page(pages[i]); 1112 put_page(pages[i]); 1113 } 1114 mem_cgroup_uncharge_end(); 1115 kfree(pages); 1116 goto out; 1117 } 1118 1119 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma, 1120 unsigned long address, pmd_t *pmd, pmd_t orig_pmd) 1121 { 1122 spinlock_t *ptl; 1123 int ret = 0; 1124 struct page *page = NULL, *new_page; 1125 unsigned long haddr; 1126 unsigned long mmun_start; /* For mmu_notifiers */ 1127 unsigned long mmun_end; /* For mmu_notifiers */ 1128 1129 ptl = pmd_lockptr(mm, pmd); 1130 VM_BUG_ON(!vma->anon_vma); 1131 haddr = address & HPAGE_PMD_MASK; 1132 if (is_huge_zero_pmd(orig_pmd)) 1133 goto alloc; 1134 spin_lock(ptl); 1135 if (unlikely(!pmd_same(*pmd, orig_pmd))) 1136 goto out_unlock; 1137 1138 page = pmd_page(orig_pmd); 1139 VM_BUG_ON(!PageCompound(page) || !PageHead(page)); 1140 if (page_mapcount(page) == 1) { 1141 pmd_t entry; 1142 entry = pmd_mkyoung(orig_pmd); 1143 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 1144 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1)) 1145 update_mmu_cache_pmd(vma, address, pmd); 1146 ret |= VM_FAULT_WRITE; 1147 goto out_unlock; 1148 } 1149 get_page(page); 1150 spin_unlock(ptl); 1151 alloc: 1152 if (transparent_hugepage_enabled(vma) && 1153 !transparent_hugepage_debug_cow()) 1154 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma), 1155 vma, haddr, numa_node_id(), 0); 1156 else 1157 new_page = NULL; 1158 1159 if (unlikely(!new_page)) { 1160 if (is_huge_zero_pmd(orig_pmd)) { 1161 ret = do_huge_pmd_wp_zero_page_fallback(mm, vma, 1162 address, pmd, orig_pmd, haddr); 1163 } else { 1164 ret = do_huge_pmd_wp_page_fallback(mm, vma, address, 1165 pmd, orig_pmd, page, haddr); 1166 if (ret & VM_FAULT_OOM) 1167 split_huge_page(page); 1168 put_page(page); 1169 } 1170 count_vm_event(THP_FAULT_FALLBACK); 1171 goto out; 1172 } 1173 1174 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) { 1175 put_page(new_page); 1176 if (page) { 1177 split_huge_page(page); 1178 put_page(page); 1179 } 1180 count_vm_event(THP_FAULT_FALLBACK); 1181 ret |= VM_FAULT_OOM; 1182 goto out; 1183 } 1184 1185 count_vm_event(THP_FAULT_ALLOC); 1186 1187 if (is_huge_zero_pmd(orig_pmd)) 1188 clear_huge_page(new_page, haddr, HPAGE_PMD_NR); 1189 else 1190 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR); 1191 __SetPageUptodate(new_page); 1192 1193 mmun_start = haddr; 1194 mmun_end = haddr + HPAGE_PMD_SIZE; 1195 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 1196 1197 spin_lock(ptl); 1198 if (page) 1199 put_page(page); 1200 if (unlikely(!pmd_same(*pmd, orig_pmd))) { 1201 spin_unlock(ptl); 1202 mem_cgroup_uncharge_page(new_page); 1203 put_page(new_page); 1204 goto out_mn; 1205 } else { 1206 pmd_t entry; 1207 entry = mk_huge_pmd(new_page, vma->vm_page_prot); 1208 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 1209 pmdp_clear_flush(vma, haddr, pmd); 1210 page_add_new_anon_rmap(new_page, vma, haddr); 1211 set_pmd_at(mm, haddr, pmd, entry); 1212 update_mmu_cache_pmd(vma, address, pmd); 1213 if (is_huge_zero_pmd(orig_pmd)) { 1214 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR); 1215 put_huge_zero_page(); 1216 } else { 1217 VM_BUG_ON(!PageHead(page)); 1218 page_remove_rmap(page); 1219 put_page(page); 1220 } 1221 ret |= VM_FAULT_WRITE; 1222 } 1223 spin_unlock(ptl); 1224 out_mn: 1225 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 1226 out: 1227 return ret; 1228 out_unlock: 1229 spin_unlock(ptl); 1230 return ret; 1231 } 1232 1233 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, 1234 unsigned long addr, 1235 pmd_t *pmd, 1236 unsigned int flags) 1237 { 1238 struct mm_struct *mm = vma->vm_mm; 1239 struct page *page = NULL; 1240 1241 assert_spin_locked(pmd_lockptr(mm, pmd)); 1242 1243 if (flags & FOLL_WRITE && !pmd_write(*pmd)) 1244 goto out; 1245 1246 /* Avoid dumping huge zero page */ 1247 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd)) 1248 return ERR_PTR(-EFAULT); 1249 1250 /* Full NUMA hinting faults to serialise migration in fault paths */ 1251 if ((flags & FOLL_NUMA) && pmd_numa(*pmd)) 1252 goto out; 1253 1254 page = pmd_page(*pmd); 1255 VM_BUG_ON(!PageHead(page)); 1256 if (flags & FOLL_TOUCH) { 1257 pmd_t _pmd; 1258 /* 1259 * We should set the dirty bit only for FOLL_WRITE but 1260 * for now the dirty bit in the pmd is meaningless. 1261 * And if the dirty bit will become meaningful and 1262 * we'll only set it with FOLL_WRITE, an atomic 1263 * set_bit will be required on the pmd to set the 1264 * young bit, instead of the current set_pmd_at. 1265 */ 1266 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd)); 1267 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK, 1268 pmd, _pmd, 1)) 1269 update_mmu_cache_pmd(vma, addr, pmd); 1270 } 1271 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) { 1272 if (page->mapping && trylock_page(page)) { 1273 lru_add_drain(); 1274 if (page->mapping) 1275 mlock_vma_page(page); 1276 unlock_page(page); 1277 } 1278 } 1279 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; 1280 VM_BUG_ON(!PageCompound(page)); 1281 if (flags & FOLL_GET) 1282 get_page_foll(page); 1283 1284 out: 1285 return page; 1286 } 1287 1288 /* NUMA hinting page fault entry point for trans huge pmds */ 1289 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma, 1290 unsigned long addr, pmd_t pmd, pmd_t *pmdp) 1291 { 1292 spinlock_t *ptl; 1293 struct anon_vma *anon_vma = NULL; 1294 struct page *page; 1295 unsigned long haddr = addr & HPAGE_PMD_MASK; 1296 int page_nid = -1, this_nid = numa_node_id(); 1297 int target_nid, last_cpupid = -1; 1298 bool page_locked; 1299 bool migrated = false; 1300 int flags = 0; 1301 1302 ptl = pmd_lock(mm, pmdp); 1303 if (unlikely(!pmd_same(pmd, *pmdp))) 1304 goto out_unlock; 1305 1306 /* 1307 * If there are potential migrations, wait for completion and retry 1308 * without disrupting NUMA hinting information. Do not relock and 1309 * check_same as the page may no longer be mapped. 1310 */ 1311 if (unlikely(pmd_trans_migrating(*pmdp))) { 1312 spin_unlock(ptl); 1313 wait_migrate_huge_page(vma->anon_vma, pmdp); 1314 goto out; 1315 } 1316 1317 page = pmd_page(pmd); 1318 BUG_ON(is_huge_zero_page(page)); 1319 page_nid = page_to_nid(page); 1320 last_cpupid = page_cpupid_last(page); 1321 count_vm_numa_event(NUMA_HINT_FAULTS); 1322 if (page_nid == this_nid) { 1323 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); 1324 flags |= TNF_FAULT_LOCAL; 1325 } 1326 1327 /* 1328 * Avoid grouping on DSO/COW pages in specific and RO pages 1329 * in general, RO pages shouldn't hurt as much anyway since 1330 * they can be in shared cache state. 1331 */ 1332 if (!pmd_write(pmd)) 1333 flags |= TNF_NO_GROUP; 1334 1335 /* 1336 * Acquire the page lock to serialise THP migrations but avoid dropping 1337 * page_table_lock if at all possible 1338 */ 1339 page_locked = trylock_page(page); 1340 target_nid = mpol_misplaced(page, vma, haddr); 1341 if (target_nid == -1) { 1342 /* If the page was locked, there are no parallel migrations */ 1343 if (page_locked) 1344 goto clear_pmdnuma; 1345 } 1346 1347 /* Migration could have started since the pmd_trans_migrating check */ 1348 if (!page_locked) { 1349 spin_unlock(ptl); 1350 wait_on_page_locked(page); 1351 page_nid = -1; 1352 goto out; 1353 } 1354 1355 /* 1356 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma 1357 * to serialises splits 1358 */ 1359 get_page(page); 1360 spin_unlock(ptl); 1361 anon_vma = page_lock_anon_vma_read(page); 1362 1363 /* Confirm the PMD did not change while page_table_lock was released */ 1364 spin_lock(ptl); 1365 if (unlikely(!pmd_same(pmd, *pmdp))) { 1366 unlock_page(page); 1367 put_page(page); 1368 page_nid = -1; 1369 goto out_unlock; 1370 } 1371 1372 /* Bail if we fail to protect against THP splits for any reason */ 1373 if (unlikely(!anon_vma)) { 1374 put_page(page); 1375 page_nid = -1; 1376 goto clear_pmdnuma; 1377 } 1378 1379 /* 1380 * Migrate the THP to the requested node, returns with page unlocked 1381 * and pmd_numa cleared. 1382 */ 1383 spin_unlock(ptl); 1384 migrated = migrate_misplaced_transhuge_page(mm, vma, 1385 pmdp, pmd, addr, page, target_nid); 1386 if (migrated) { 1387 flags |= TNF_MIGRATED; 1388 page_nid = target_nid; 1389 } 1390 1391 goto out; 1392 clear_pmdnuma: 1393 BUG_ON(!PageLocked(page)); 1394 pmd = pmd_mknonnuma(pmd); 1395 set_pmd_at(mm, haddr, pmdp, pmd); 1396 VM_BUG_ON(pmd_numa(*pmdp)); 1397 update_mmu_cache_pmd(vma, addr, pmdp); 1398 unlock_page(page); 1399 out_unlock: 1400 spin_unlock(ptl); 1401 1402 out: 1403 if (anon_vma) 1404 page_unlock_anon_vma_read(anon_vma); 1405 1406 if (page_nid != -1) 1407 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags); 1408 1409 return 0; 1410 } 1411 1412 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1413 pmd_t *pmd, unsigned long addr) 1414 { 1415 spinlock_t *ptl; 1416 int ret = 0; 1417 1418 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) { 1419 struct page *page; 1420 pgtable_t pgtable; 1421 pmd_t orig_pmd; 1422 /* 1423 * For architectures like ppc64 we look at deposited pgtable 1424 * when calling pmdp_get_and_clear. So do the 1425 * pgtable_trans_huge_withdraw after finishing pmdp related 1426 * operations. 1427 */ 1428 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd); 1429 tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 1430 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd); 1431 if (is_huge_zero_pmd(orig_pmd)) { 1432 atomic_long_dec(&tlb->mm->nr_ptes); 1433 spin_unlock(ptl); 1434 put_huge_zero_page(); 1435 } else { 1436 page = pmd_page(orig_pmd); 1437 page_remove_rmap(page); 1438 VM_BUG_ON(page_mapcount(page) < 0); 1439 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); 1440 VM_BUG_ON(!PageHead(page)); 1441 atomic_long_dec(&tlb->mm->nr_ptes); 1442 spin_unlock(ptl); 1443 tlb_remove_page(tlb, page); 1444 } 1445 pte_free(tlb->mm, pgtable); 1446 ret = 1; 1447 } 1448 return ret; 1449 } 1450 1451 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 1452 unsigned long addr, unsigned long end, 1453 unsigned char *vec) 1454 { 1455 spinlock_t *ptl; 1456 int ret = 0; 1457 1458 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) { 1459 /* 1460 * All logical pages in the range are present 1461 * if backed by a huge page. 1462 */ 1463 spin_unlock(ptl); 1464 memset(vec, 1, (end - addr) >> PAGE_SHIFT); 1465 ret = 1; 1466 } 1467 1468 return ret; 1469 } 1470 1471 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma, 1472 unsigned long old_addr, 1473 unsigned long new_addr, unsigned long old_end, 1474 pmd_t *old_pmd, pmd_t *new_pmd) 1475 { 1476 spinlock_t *old_ptl, *new_ptl; 1477 int ret = 0; 1478 pmd_t pmd; 1479 1480 struct mm_struct *mm = vma->vm_mm; 1481 1482 if ((old_addr & ~HPAGE_PMD_MASK) || 1483 (new_addr & ~HPAGE_PMD_MASK) || 1484 old_end - old_addr < HPAGE_PMD_SIZE || 1485 (new_vma->vm_flags & VM_NOHUGEPAGE)) 1486 goto out; 1487 1488 /* 1489 * The destination pmd shouldn't be established, free_pgtables() 1490 * should have release it. 1491 */ 1492 if (WARN_ON(!pmd_none(*new_pmd))) { 1493 VM_BUG_ON(pmd_trans_huge(*new_pmd)); 1494 goto out; 1495 } 1496 1497 /* 1498 * We don't have to worry about the ordering of src and dst 1499 * ptlocks because exclusive mmap_sem prevents deadlock. 1500 */ 1501 ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl); 1502 if (ret == 1) { 1503 new_ptl = pmd_lockptr(mm, new_pmd); 1504 if (new_ptl != old_ptl) 1505 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); 1506 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd); 1507 VM_BUG_ON(!pmd_none(*new_pmd)); 1508 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd)); 1509 if (new_ptl != old_ptl) { 1510 pgtable_t pgtable; 1511 1512 /* 1513 * Move preallocated PTE page table if new_pmd is on 1514 * different PMD page table. 1515 */ 1516 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd); 1517 pgtable_trans_huge_deposit(mm, new_pmd, pgtable); 1518 1519 spin_unlock(new_ptl); 1520 } 1521 spin_unlock(old_ptl); 1522 } 1523 out: 1524 return ret; 1525 } 1526 1527 /* 1528 * Returns 1529 * - 0 if PMD could not be locked 1530 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary 1531 * - HPAGE_PMD_NR is protections changed and TLB flush necessary 1532 */ 1533 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 1534 unsigned long addr, pgprot_t newprot, int prot_numa) 1535 { 1536 struct mm_struct *mm = vma->vm_mm; 1537 spinlock_t *ptl; 1538 int ret = 0; 1539 1540 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) { 1541 pmd_t entry; 1542 ret = 1; 1543 if (!prot_numa) { 1544 entry = pmdp_get_and_clear(mm, addr, pmd); 1545 if (pmd_numa(entry)) 1546 entry = pmd_mknonnuma(entry); 1547 entry = pmd_modify(entry, newprot); 1548 ret = HPAGE_PMD_NR; 1549 BUG_ON(pmd_write(entry)); 1550 } else { 1551 struct page *page = pmd_page(*pmd); 1552 1553 /* 1554 * Do not trap faults against the zero page. The 1555 * read-only data is likely to be read-cached on the 1556 * local CPU cache and it is less useful to know about 1557 * local vs remote hits on the zero page. 1558 */ 1559 if (!is_huge_zero_page(page) && 1560 !pmd_numa(*pmd)) { 1561 entry = *pmd; 1562 entry = pmd_mknuma(entry); 1563 ret = HPAGE_PMD_NR; 1564 } 1565 } 1566 1567 /* Set PMD if cleared earlier */ 1568 if (ret == HPAGE_PMD_NR) 1569 set_pmd_at(mm, addr, pmd, entry); 1570 1571 spin_unlock(ptl); 1572 } 1573 1574 return ret; 1575 } 1576 1577 /* 1578 * Returns 1 if a given pmd maps a stable (not under splitting) thp. 1579 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise. 1580 * 1581 * Note that if it returns 1, this routine returns without unlocking page 1582 * table locks. So callers must unlock them. 1583 */ 1584 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma, 1585 spinlock_t **ptl) 1586 { 1587 *ptl = pmd_lock(vma->vm_mm, pmd); 1588 if (likely(pmd_trans_huge(*pmd))) { 1589 if (unlikely(pmd_trans_splitting(*pmd))) { 1590 spin_unlock(*ptl); 1591 wait_split_huge_page(vma->anon_vma, pmd); 1592 return -1; 1593 } else { 1594 /* Thp mapped by 'pmd' is stable, so we can 1595 * handle it as it is. */ 1596 return 1; 1597 } 1598 } 1599 spin_unlock(*ptl); 1600 return 0; 1601 } 1602 1603 /* 1604 * This function returns whether a given @page is mapped onto the @address 1605 * in the virtual space of @mm. 1606 * 1607 * When it's true, this function returns *pmd with holding the page table lock 1608 * and passing it back to the caller via @ptl. 1609 * If it's false, returns NULL without holding the page table lock. 1610 */ 1611 pmd_t *page_check_address_pmd(struct page *page, 1612 struct mm_struct *mm, 1613 unsigned long address, 1614 enum page_check_address_pmd_flag flag, 1615 spinlock_t **ptl) 1616 { 1617 pmd_t *pmd; 1618 1619 if (address & ~HPAGE_PMD_MASK) 1620 return NULL; 1621 1622 pmd = mm_find_pmd(mm, address); 1623 if (!pmd) 1624 return NULL; 1625 *ptl = pmd_lock(mm, pmd); 1626 if (pmd_none(*pmd)) 1627 goto unlock; 1628 if (pmd_page(*pmd) != page) 1629 goto unlock; 1630 /* 1631 * split_vma() may create temporary aliased mappings. There is 1632 * no risk as long as all huge pmd are found and have their 1633 * splitting bit set before __split_huge_page_refcount 1634 * runs. Finding the same huge pmd more than once during the 1635 * same rmap walk is not a problem. 1636 */ 1637 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG && 1638 pmd_trans_splitting(*pmd)) 1639 goto unlock; 1640 if (pmd_trans_huge(*pmd)) { 1641 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG && 1642 !pmd_trans_splitting(*pmd)); 1643 return pmd; 1644 } 1645 unlock: 1646 spin_unlock(*ptl); 1647 return NULL; 1648 } 1649 1650 static int __split_huge_page_splitting(struct page *page, 1651 struct vm_area_struct *vma, 1652 unsigned long address) 1653 { 1654 struct mm_struct *mm = vma->vm_mm; 1655 spinlock_t *ptl; 1656 pmd_t *pmd; 1657 int ret = 0; 1658 /* For mmu_notifiers */ 1659 const unsigned long mmun_start = address; 1660 const unsigned long mmun_end = address + HPAGE_PMD_SIZE; 1661 1662 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 1663 pmd = page_check_address_pmd(page, mm, address, 1664 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl); 1665 if (pmd) { 1666 /* 1667 * We can't temporarily set the pmd to null in order 1668 * to split it, the pmd must remain marked huge at all 1669 * times or the VM won't take the pmd_trans_huge paths 1670 * and it won't wait on the anon_vma->root->rwsem to 1671 * serialize against split_huge_page*. 1672 */ 1673 pmdp_splitting_flush(vma, address, pmd); 1674 ret = 1; 1675 spin_unlock(ptl); 1676 } 1677 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 1678 1679 return ret; 1680 } 1681 1682 static void __split_huge_page_refcount(struct page *page, 1683 struct list_head *list) 1684 { 1685 int i; 1686 struct zone *zone = page_zone(page); 1687 struct lruvec *lruvec; 1688 int tail_count = 0; 1689 1690 /* prevent PageLRU to go away from under us, and freeze lru stats */ 1691 spin_lock_irq(&zone->lru_lock); 1692 lruvec = mem_cgroup_page_lruvec(page, zone); 1693 1694 compound_lock(page); 1695 /* complete memcg works before add pages to LRU */ 1696 mem_cgroup_split_huge_fixup(page); 1697 1698 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) { 1699 struct page *page_tail = page + i; 1700 1701 /* tail_page->_mapcount cannot change */ 1702 BUG_ON(page_mapcount(page_tail) < 0); 1703 tail_count += page_mapcount(page_tail); 1704 /* check for overflow */ 1705 BUG_ON(tail_count < 0); 1706 BUG_ON(atomic_read(&page_tail->_count) != 0); 1707 /* 1708 * tail_page->_count is zero and not changing from 1709 * under us. But get_page_unless_zero() may be running 1710 * from under us on the tail_page. If we used 1711 * atomic_set() below instead of atomic_add(), we 1712 * would then run atomic_set() concurrently with 1713 * get_page_unless_zero(), and atomic_set() is 1714 * implemented in C not using locked ops. spin_unlock 1715 * on x86 sometime uses locked ops because of PPro 1716 * errata 66, 92, so unless somebody can guarantee 1717 * atomic_set() here would be safe on all archs (and 1718 * not only on x86), it's safer to use atomic_add(). 1719 */ 1720 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1, 1721 &page_tail->_count); 1722 1723 /* after clearing PageTail the gup refcount can be released */ 1724 smp_mb(); 1725 1726 /* 1727 * retain hwpoison flag of the poisoned tail page: 1728 * fix for the unsuitable process killed on Guest Machine(KVM) 1729 * by the memory-failure. 1730 */ 1731 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON; 1732 page_tail->flags |= (page->flags & 1733 ((1L << PG_referenced) | 1734 (1L << PG_swapbacked) | 1735 (1L << PG_mlocked) | 1736 (1L << PG_uptodate) | 1737 (1L << PG_active) | 1738 (1L << PG_unevictable))); 1739 page_tail->flags |= (1L << PG_dirty); 1740 1741 /* clear PageTail before overwriting first_page */ 1742 smp_wmb(); 1743 1744 /* 1745 * __split_huge_page_splitting() already set the 1746 * splitting bit in all pmd that could map this 1747 * hugepage, that will ensure no CPU can alter the 1748 * mapcount on the head page. The mapcount is only 1749 * accounted in the head page and it has to be 1750 * transferred to all tail pages in the below code. So 1751 * for this code to be safe, the split the mapcount 1752 * can't change. But that doesn't mean userland can't 1753 * keep changing and reading the page contents while 1754 * we transfer the mapcount, so the pmd splitting 1755 * status is achieved setting a reserved bit in the 1756 * pmd, not by clearing the present bit. 1757 */ 1758 page_tail->_mapcount = page->_mapcount; 1759 1760 BUG_ON(page_tail->mapping); 1761 page_tail->mapping = page->mapping; 1762 1763 page_tail->index = page->index + i; 1764 page_cpupid_xchg_last(page_tail, page_cpupid_last(page)); 1765 1766 BUG_ON(!PageAnon(page_tail)); 1767 BUG_ON(!PageUptodate(page_tail)); 1768 BUG_ON(!PageDirty(page_tail)); 1769 BUG_ON(!PageSwapBacked(page_tail)); 1770 1771 lru_add_page_tail(page, page_tail, lruvec, list); 1772 } 1773 atomic_sub(tail_count, &page->_count); 1774 BUG_ON(atomic_read(&page->_count) <= 0); 1775 1776 __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1); 1777 1778 ClearPageCompound(page); 1779 compound_unlock(page); 1780 spin_unlock_irq(&zone->lru_lock); 1781 1782 for (i = 1; i < HPAGE_PMD_NR; i++) { 1783 struct page *page_tail = page + i; 1784 BUG_ON(page_count(page_tail) <= 0); 1785 /* 1786 * Tail pages may be freed if there wasn't any mapping 1787 * like if add_to_swap() is running on a lru page that 1788 * had its mapping zapped. And freeing these pages 1789 * requires taking the lru_lock so we do the put_page 1790 * of the tail pages after the split is complete. 1791 */ 1792 put_page(page_tail); 1793 } 1794 1795 /* 1796 * Only the head page (now become a regular page) is required 1797 * to be pinned by the caller. 1798 */ 1799 BUG_ON(page_count(page) <= 0); 1800 } 1801 1802 static int __split_huge_page_map(struct page *page, 1803 struct vm_area_struct *vma, 1804 unsigned long address) 1805 { 1806 struct mm_struct *mm = vma->vm_mm; 1807 spinlock_t *ptl; 1808 pmd_t *pmd, _pmd; 1809 int ret = 0, i; 1810 pgtable_t pgtable; 1811 unsigned long haddr; 1812 1813 pmd = page_check_address_pmd(page, mm, address, 1814 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl); 1815 if (pmd) { 1816 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 1817 pmd_populate(mm, &_pmd, pgtable); 1818 1819 haddr = address; 1820 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 1821 pte_t *pte, entry; 1822 BUG_ON(PageCompound(page+i)); 1823 entry = mk_pte(page + i, vma->vm_page_prot); 1824 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 1825 if (!pmd_write(*pmd)) 1826 entry = pte_wrprotect(entry); 1827 else 1828 BUG_ON(page_mapcount(page) != 1); 1829 if (!pmd_young(*pmd)) 1830 entry = pte_mkold(entry); 1831 if (pmd_numa(*pmd)) 1832 entry = pte_mknuma(entry); 1833 pte = pte_offset_map(&_pmd, haddr); 1834 BUG_ON(!pte_none(*pte)); 1835 set_pte_at(mm, haddr, pte, entry); 1836 pte_unmap(pte); 1837 } 1838 1839 smp_wmb(); /* make pte visible before pmd */ 1840 /* 1841 * Up to this point the pmd is present and huge and 1842 * userland has the whole access to the hugepage 1843 * during the split (which happens in place). If we 1844 * overwrite the pmd with the not-huge version 1845 * pointing to the pte here (which of course we could 1846 * if all CPUs were bug free), userland could trigger 1847 * a small page size TLB miss on the small sized TLB 1848 * while the hugepage TLB entry is still established 1849 * in the huge TLB. Some CPU doesn't like that. See 1850 * http://support.amd.com/us/Processor_TechDocs/41322.pdf, 1851 * Erratum 383 on page 93. Intel should be safe but is 1852 * also warns that it's only safe if the permission 1853 * and cache attributes of the two entries loaded in 1854 * the two TLB is identical (which should be the case 1855 * here). But it is generally safer to never allow 1856 * small and huge TLB entries for the same virtual 1857 * address to be loaded simultaneously. So instead of 1858 * doing "pmd_populate(); flush_tlb_range();" we first 1859 * mark the current pmd notpresent (atomically because 1860 * here the pmd_trans_huge and pmd_trans_splitting 1861 * must remain set at all times on the pmd until the 1862 * split is complete for this pmd), then we flush the 1863 * SMP TLB and finally we write the non-huge version 1864 * of the pmd entry with pmd_populate. 1865 */ 1866 pmdp_invalidate(vma, address, pmd); 1867 pmd_populate(mm, pmd, pgtable); 1868 ret = 1; 1869 spin_unlock(ptl); 1870 } 1871 1872 return ret; 1873 } 1874 1875 /* must be called with anon_vma->root->rwsem held */ 1876 static void __split_huge_page(struct page *page, 1877 struct anon_vma *anon_vma, 1878 struct list_head *list) 1879 { 1880 int mapcount, mapcount2; 1881 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); 1882 struct anon_vma_chain *avc; 1883 1884 BUG_ON(!PageHead(page)); 1885 BUG_ON(PageTail(page)); 1886 1887 mapcount = 0; 1888 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) { 1889 struct vm_area_struct *vma = avc->vma; 1890 unsigned long addr = vma_address(page, vma); 1891 BUG_ON(is_vma_temporary_stack(vma)); 1892 mapcount += __split_huge_page_splitting(page, vma, addr); 1893 } 1894 /* 1895 * It is critical that new vmas are added to the tail of the 1896 * anon_vma list. This guarantes that if copy_huge_pmd() runs 1897 * and establishes a child pmd before 1898 * __split_huge_page_splitting() freezes the parent pmd (so if 1899 * we fail to prevent copy_huge_pmd() from running until the 1900 * whole __split_huge_page() is complete), we will still see 1901 * the newly established pmd of the child later during the 1902 * walk, to be able to set it as pmd_trans_splitting too. 1903 */ 1904 if (mapcount != page_mapcount(page)) 1905 printk(KERN_ERR "mapcount %d page_mapcount %d\n", 1906 mapcount, page_mapcount(page)); 1907 BUG_ON(mapcount != page_mapcount(page)); 1908 1909 __split_huge_page_refcount(page, list); 1910 1911 mapcount2 = 0; 1912 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) { 1913 struct vm_area_struct *vma = avc->vma; 1914 unsigned long addr = vma_address(page, vma); 1915 BUG_ON(is_vma_temporary_stack(vma)); 1916 mapcount2 += __split_huge_page_map(page, vma, addr); 1917 } 1918 if (mapcount != mapcount2) 1919 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n", 1920 mapcount, mapcount2, page_mapcount(page)); 1921 BUG_ON(mapcount != mapcount2); 1922 } 1923 1924 /* 1925 * Split a hugepage into normal pages. This doesn't change the position of head 1926 * page. If @list is null, tail pages will be added to LRU list, otherwise, to 1927 * @list. Both head page and tail pages will inherit mapping, flags, and so on 1928 * from the hugepage. 1929 * Return 0 if the hugepage is split successfully otherwise return 1. 1930 */ 1931 int split_huge_page_to_list(struct page *page, struct list_head *list) 1932 { 1933 struct anon_vma *anon_vma; 1934 int ret = 1; 1935 1936 BUG_ON(is_huge_zero_page(page)); 1937 BUG_ON(!PageAnon(page)); 1938 1939 /* 1940 * The caller does not necessarily hold an mmap_sem that would prevent 1941 * the anon_vma disappearing so we first we take a reference to it 1942 * and then lock the anon_vma for write. This is similar to 1943 * page_lock_anon_vma_read except the write lock is taken to serialise 1944 * against parallel split or collapse operations. 1945 */ 1946 anon_vma = page_get_anon_vma(page); 1947 if (!anon_vma) 1948 goto out; 1949 anon_vma_lock_write(anon_vma); 1950 1951 ret = 0; 1952 if (!PageCompound(page)) 1953 goto out_unlock; 1954 1955 BUG_ON(!PageSwapBacked(page)); 1956 __split_huge_page(page, anon_vma, list); 1957 count_vm_event(THP_SPLIT); 1958 1959 BUG_ON(PageCompound(page)); 1960 out_unlock: 1961 anon_vma_unlock_write(anon_vma); 1962 put_anon_vma(anon_vma); 1963 out: 1964 return ret; 1965 } 1966 1967 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE) 1968 1969 int hugepage_madvise(struct vm_area_struct *vma, 1970 unsigned long *vm_flags, int advice) 1971 { 1972 struct mm_struct *mm = vma->vm_mm; 1973 1974 switch (advice) { 1975 case MADV_HUGEPAGE: 1976 /* 1977 * Be somewhat over-protective like KSM for now! 1978 */ 1979 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP)) 1980 return -EINVAL; 1981 if (mm->def_flags & VM_NOHUGEPAGE) 1982 return -EINVAL; 1983 *vm_flags &= ~VM_NOHUGEPAGE; 1984 *vm_flags |= VM_HUGEPAGE; 1985 /* 1986 * If the vma become good for khugepaged to scan, 1987 * register it here without waiting a page fault that 1988 * may not happen any time soon. 1989 */ 1990 if (unlikely(khugepaged_enter_vma_merge(vma))) 1991 return -ENOMEM; 1992 break; 1993 case MADV_NOHUGEPAGE: 1994 /* 1995 * Be somewhat over-protective like KSM for now! 1996 */ 1997 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP)) 1998 return -EINVAL; 1999 *vm_flags &= ~VM_HUGEPAGE; 2000 *vm_flags |= VM_NOHUGEPAGE; 2001 /* 2002 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning 2003 * this vma even if we leave the mm registered in khugepaged if 2004 * it got registered before VM_NOHUGEPAGE was set. 2005 */ 2006 break; 2007 } 2008 2009 return 0; 2010 } 2011 2012 static int __init khugepaged_slab_init(void) 2013 { 2014 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot", 2015 sizeof(struct mm_slot), 2016 __alignof__(struct mm_slot), 0, NULL); 2017 if (!mm_slot_cache) 2018 return -ENOMEM; 2019 2020 return 0; 2021 } 2022 2023 static inline struct mm_slot *alloc_mm_slot(void) 2024 { 2025 if (!mm_slot_cache) /* initialization failed */ 2026 return NULL; 2027 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL); 2028 } 2029 2030 static inline void free_mm_slot(struct mm_slot *mm_slot) 2031 { 2032 kmem_cache_free(mm_slot_cache, mm_slot); 2033 } 2034 2035 static struct mm_slot *get_mm_slot(struct mm_struct *mm) 2036 { 2037 struct mm_slot *mm_slot; 2038 2039 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm) 2040 if (mm == mm_slot->mm) 2041 return mm_slot; 2042 2043 return NULL; 2044 } 2045 2046 static void insert_to_mm_slots_hash(struct mm_struct *mm, 2047 struct mm_slot *mm_slot) 2048 { 2049 mm_slot->mm = mm; 2050 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm); 2051 } 2052 2053 static inline int khugepaged_test_exit(struct mm_struct *mm) 2054 { 2055 return atomic_read(&mm->mm_users) == 0; 2056 } 2057 2058 int __khugepaged_enter(struct mm_struct *mm) 2059 { 2060 struct mm_slot *mm_slot; 2061 int wakeup; 2062 2063 mm_slot = alloc_mm_slot(); 2064 if (!mm_slot) 2065 return -ENOMEM; 2066 2067 /* __khugepaged_exit() must not run from under us */ 2068 VM_BUG_ON(khugepaged_test_exit(mm)); 2069 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) { 2070 free_mm_slot(mm_slot); 2071 return 0; 2072 } 2073 2074 spin_lock(&khugepaged_mm_lock); 2075 insert_to_mm_slots_hash(mm, mm_slot); 2076 /* 2077 * Insert just behind the scanning cursor, to let the area settle 2078 * down a little. 2079 */ 2080 wakeup = list_empty(&khugepaged_scan.mm_head); 2081 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head); 2082 spin_unlock(&khugepaged_mm_lock); 2083 2084 atomic_inc(&mm->mm_count); 2085 if (wakeup) 2086 wake_up_interruptible(&khugepaged_wait); 2087 2088 return 0; 2089 } 2090 2091 int khugepaged_enter_vma_merge(struct vm_area_struct *vma) 2092 { 2093 unsigned long hstart, hend; 2094 if (!vma->anon_vma) 2095 /* 2096 * Not yet faulted in so we will register later in the 2097 * page fault if needed. 2098 */ 2099 return 0; 2100 if (vma->vm_ops) 2101 /* khugepaged not yet working on file or special mappings */ 2102 return 0; 2103 VM_BUG_ON(vma->vm_flags & VM_NO_THP); 2104 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; 2105 hend = vma->vm_end & HPAGE_PMD_MASK; 2106 if (hstart < hend) 2107 return khugepaged_enter(vma); 2108 return 0; 2109 } 2110 2111 void __khugepaged_exit(struct mm_struct *mm) 2112 { 2113 struct mm_slot *mm_slot; 2114 int free = 0; 2115 2116 spin_lock(&khugepaged_mm_lock); 2117 mm_slot = get_mm_slot(mm); 2118 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) { 2119 hash_del(&mm_slot->hash); 2120 list_del(&mm_slot->mm_node); 2121 free = 1; 2122 } 2123 spin_unlock(&khugepaged_mm_lock); 2124 2125 if (free) { 2126 clear_bit(MMF_VM_HUGEPAGE, &mm->flags); 2127 free_mm_slot(mm_slot); 2128 mmdrop(mm); 2129 } else if (mm_slot) { 2130 /* 2131 * This is required to serialize against 2132 * khugepaged_test_exit() (which is guaranteed to run 2133 * under mmap sem read mode). Stop here (after we 2134 * return all pagetables will be destroyed) until 2135 * khugepaged has finished working on the pagetables 2136 * under the mmap_sem. 2137 */ 2138 down_write(&mm->mmap_sem); 2139 up_write(&mm->mmap_sem); 2140 } 2141 } 2142 2143 static void release_pte_page(struct page *page) 2144 { 2145 /* 0 stands for page_is_file_cache(page) == false */ 2146 dec_zone_page_state(page, NR_ISOLATED_ANON + 0); 2147 unlock_page(page); 2148 putback_lru_page(page); 2149 } 2150 2151 static void release_pte_pages(pte_t *pte, pte_t *_pte) 2152 { 2153 while (--_pte >= pte) { 2154 pte_t pteval = *_pte; 2155 if (!pte_none(pteval)) 2156 release_pte_page(pte_page(pteval)); 2157 } 2158 } 2159 2160 static int __collapse_huge_page_isolate(struct vm_area_struct *vma, 2161 unsigned long address, 2162 pte_t *pte) 2163 { 2164 struct page *page; 2165 pte_t *_pte; 2166 int referenced = 0, none = 0; 2167 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; 2168 _pte++, address += PAGE_SIZE) { 2169 pte_t pteval = *_pte; 2170 if (pte_none(pteval)) { 2171 if (++none <= khugepaged_max_ptes_none) 2172 continue; 2173 else 2174 goto out; 2175 } 2176 if (!pte_present(pteval) || !pte_write(pteval)) 2177 goto out; 2178 page = vm_normal_page(vma, address, pteval); 2179 if (unlikely(!page)) 2180 goto out; 2181 2182 VM_BUG_ON(PageCompound(page)); 2183 BUG_ON(!PageAnon(page)); 2184 VM_BUG_ON(!PageSwapBacked(page)); 2185 2186 /* cannot use mapcount: can't collapse if there's a gup pin */ 2187 if (page_count(page) != 1) 2188 goto out; 2189 /* 2190 * We can do it before isolate_lru_page because the 2191 * page can't be freed from under us. NOTE: PG_lock 2192 * is needed to serialize against split_huge_page 2193 * when invoked from the VM. 2194 */ 2195 if (!trylock_page(page)) 2196 goto out; 2197 /* 2198 * Isolate the page to avoid collapsing an hugepage 2199 * currently in use by the VM. 2200 */ 2201 if (isolate_lru_page(page)) { 2202 unlock_page(page); 2203 goto out; 2204 } 2205 /* 0 stands for page_is_file_cache(page) == false */ 2206 inc_zone_page_state(page, NR_ISOLATED_ANON + 0); 2207 VM_BUG_ON(!PageLocked(page)); 2208 VM_BUG_ON(PageLRU(page)); 2209 2210 /* If there is no mapped pte young don't collapse the page */ 2211 if (pte_young(pteval) || PageReferenced(page) || 2212 mmu_notifier_test_young(vma->vm_mm, address)) 2213 referenced = 1; 2214 } 2215 if (likely(referenced)) 2216 return 1; 2217 out: 2218 release_pte_pages(pte, _pte); 2219 return 0; 2220 } 2221 2222 static void __collapse_huge_page_copy(pte_t *pte, struct page *page, 2223 struct vm_area_struct *vma, 2224 unsigned long address, 2225 spinlock_t *ptl) 2226 { 2227 pte_t *_pte; 2228 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) { 2229 pte_t pteval = *_pte; 2230 struct page *src_page; 2231 2232 if (pte_none(pteval)) { 2233 clear_user_highpage(page, address); 2234 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); 2235 } else { 2236 src_page = pte_page(pteval); 2237 copy_user_highpage(page, src_page, address, vma); 2238 VM_BUG_ON(page_mapcount(src_page) != 1); 2239 release_pte_page(src_page); 2240 /* 2241 * ptl mostly unnecessary, but preempt has to 2242 * be disabled to update the per-cpu stats 2243 * inside page_remove_rmap(). 2244 */ 2245 spin_lock(ptl); 2246 /* 2247 * paravirt calls inside pte_clear here are 2248 * superfluous. 2249 */ 2250 pte_clear(vma->vm_mm, address, _pte); 2251 page_remove_rmap(src_page); 2252 spin_unlock(ptl); 2253 free_page_and_swap_cache(src_page); 2254 } 2255 2256 address += PAGE_SIZE; 2257 page++; 2258 } 2259 } 2260 2261 static void khugepaged_alloc_sleep(void) 2262 { 2263 wait_event_freezable_timeout(khugepaged_wait, false, 2264 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs)); 2265 } 2266 2267 static int khugepaged_node_load[MAX_NUMNODES]; 2268 2269 #ifdef CONFIG_NUMA 2270 static int khugepaged_find_target_node(void) 2271 { 2272 static int last_khugepaged_target_node = NUMA_NO_NODE; 2273 int nid, target_node = 0, max_value = 0; 2274 2275 /* find first node with max normal pages hit */ 2276 for (nid = 0; nid < MAX_NUMNODES; nid++) 2277 if (khugepaged_node_load[nid] > max_value) { 2278 max_value = khugepaged_node_load[nid]; 2279 target_node = nid; 2280 } 2281 2282 /* do some balance if several nodes have the same hit record */ 2283 if (target_node <= last_khugepaged_target_node) 2284 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES; 2285 nid++) 2286 if (max_value == khugepaged_node_load[nid]) { 2287 target_node = nid; 2288 break; 2289 } 2290 2291 last_khugepaged_target_node = target_node; 2292 return target_node; 2293 } 2294 2295 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) 2296 { 2297 if (IS_ERR(*hpage)) { 2298 if (!*wait) 2299 return false; 2300 2301 *wait = false; 2302 *hpage = NULL; 2303 khugepaged_alloc_sleep(); 2304 } else if (*hpage) { 2305 put_page(*hpage); 2306 *hpage = NULL; 2307 } 2308 2309 return true; 2310 } 2311 2312 static struct page 2313 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm, 2314 struct vm_area_struct *vma, unsigned long address, 2315 int node) 2316 { 2317 VM_BUG_ON(*hpage); 2318 /* 2319 * Allocate the page while the vma is still valid and under 2320 * the mmap_sem read mode so there is no memory allocation 2321 * later when we take the mmap_sem in write mode. This is more 2322 * friendly behavior (OTOH it may actually hide bugs) to 2323 * filesystems in userland with daemons allocating memory in 2324 * the userland I/O paths. Allocating memory with the 2325 * mmap_sem in read mode is good idea also to allow greater 2326 * scalability. 2327 */ 2328 *hpage = alloc_pages_exact_node(node, alloc_hugepage_gfpmask( 2329 khugepaged_defrag(), __GFP_OTHER_NODE), HPAGE_PMD_ORDER); 2330 /* 2331 * After allocating the hugepage, release the mmap_sem read lock in 2332 * preparation for taking it in write mode. 2333 */ 2334 up_read(&mm->mmap_sem); 2335 if (unlikely(!*hpage)) { 2336 count_vm_event(THP_COLLAPSE_ALLOC_FAILED); 2337 *hpage = ERR_PTR(-ENOMEM); 2338 return NULL; 2339 } 2340 2341 count_vm_event(THP_COLLAPSE_ALLOC); 2342 return *hpage; 2343 } 2344 #else 2345 static int khugepaged_find_target_node(void) 2346 { 2347 return 0; 2348 } 2349 2350 static inline struct page *alloc_hugepage(int defrag) 2351 { 2352 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0), 2353 HPAGE_PMD_ORDER); 2354 } 2355 2356 static struct page *khugepaged_alloc_hugepage(bool *wait) 2357 { 2358 struct page *hpage; 2359 2360 do { 2361 hpage = alloc_hugepage(khugepaged_defrag()); 2362 if (!hpage) { 2363 count_vm_event(THP_COLLAPSE_ALLOC_FAILED); 2364 if (!*wait) 2365 return NULL; 2366 2367 *wait = false; 2368 khugepaged_alloc_sleep(); 2369 } else 2370 count_vm_event(THP_COLLAPSE_ALLOC); 2371 } while (unlikely(!hpage) && likely(khugepaged_enabled())); 2372 2373 return hpage; 2374 } 2375 2376 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) 2377 { 2378 if (!*hpage) 2379 *hpage = khugepaged_alloc_hugepage(wait); 2380 2381 if (unlikely(!*hpage)) 2382 return false; 2383 2384 return true; 2385 } 2386 2387 static struct page 2388 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm, 2389 struct vm_area_struct *vma, unsigned long address, 2390 int node) 2391 { 2392 up_read(&mm->mmap_sem); 2393 VM_BUG_ON(!*hpage); 2394 return *hpage; 2395 } 2396 #endif 2397 2398 static bool hugepage_vma_check(struct vm_area_struct *vma) 2399 { 2400 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) || 2401 (vma->vm_flags & VM_NOHUGEPAGE)) 2402 return false; 2403 2404 if (!vma->anon_vma || vma->vm_ops) 2405 return false; 2406 if (is_vma_temporary_stack(vma)) 2407 return false; 2408 VM_BUG_ON(vma->vm_flags & VM_NO_THP); 2409 return true; 2410 } 2411 2412 static void collapse_huge_page(struct mm_struct *mm, 2413 unsigned long address, 2414 struct page **hpage, 2415 struct vm_area_struct *vma, 2416 int node) 2417 { 2418 pmd_t *pmd, _pmd; 2419 pte_t *pte; 2420 pgtable_t pgtable; 2421 struct page *new_page; 2422 spinlock_t *pmd_ptl, *pte_ptl; 2423 int isolated; 2424 unsigned long hstart, hend; 2425 unsigned long mmun_start; /* For mmu_notifiers */ 2426 unsigned long mmun_end; /* For mmu_notifiers */ 2427 2428 VM_BUG_ON(address & ~HPAGE_PMD_MASK); 2429 2430 /* release the mmap_sem read lock. */ 2431 new_page = khugepaged_alloc_page(hpage, mm, vma, address, node); 2432 if (!new_page) 2433 return; 2434 2435 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) 2436 return; 2437 2438 /* 2439 * Prevent all access to pagetables with the exception of 2440 * gup_fast later hanlded by the ptep_clear_flush and the VM 2441 * handled by the anon_vma lock + PG_lock. 2442 */ 2443 down_write(&mm->mmap_sem); 2444 if (unlikely(khugepaged_test_exit(mm))) 2445 goto out; 2446 2447 vma = find_vma(mm, address); 2448 if (!vma) 2449 goto out; 2450 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; 2451 hend = vma->vm_end & HPAGE_PMD_MASK; 2452 if (address < hstart || address + HPAGE_PMD_SIZE > hend) 2453 goto out; 2454 if (!hugepage_vma_check(vma)) 2455 goto out; 2456 pmd = mm_find_pmd(mm, address); 2457 if (!pmd) 2458 goto out; 2459 if (pmd_trans_huge(*pmd)) 2460 goto out; 2461 2462 anon_vma_lock_write(vma->anon_vma); 2463 2464 pte = pte_offset_map(pmd, address); 2465 pte_ptl = pte_lockptr(mm, pmd); 2466 2467 mmun_start = address; 2468 mmun_end = address + HPAGE_PMD_SIZE; 2469 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 2470 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */ 2471 /* 2472 * After this gup_fast can't run anymore. This also removes 2473 * any huge TLB entry from the CPU so we won't allow 2474 * huge and small TLB entries for the same virtual address 2475 * to avoid the risk of CPU bugs in that area. 2476 */ 2477 _pmd = pmdp_clear_flush(vma, address, pmd); 2478 spin_unlock(pmd_ptl); 2479 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 2480 2481 spin_lock(pte_ptl); 2482 isolated = __collapse_huge_page_isolate(vma, address, pte); 2483 spin_unlock(pte_ptl); 2484 2485 if (unlikely(!isolated)) { 2486 pte_unmap(pte); 2487 spin_lock(pmd_ptl); 2488 BUG_ON(!pmd_none(*pmd)); 2489 /* 2490 * We can only use set_pmd_at when establishing 2491 * hugepmds and never for establishing regular pmds that 2492 * points to regular pagetables. Use pmd_populate for that 2493 */ 2494 pmd_populate(mm, pmd, pmd_pgtable(_pmd)); 2495 spin_unlock(pmd_ptl); 2496 anon_vma_unlock_write(vma->anon_vma); 2497 goto out; 2498 } 2499 2500 /* 2501 * All pages are isolated and locked so anon_vma rmap 2502 * can't run anymore. 2503 */ 2504 anon_vma_unlock_write(vma->anon_vma); 2505 2506 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl); 2507 pte_unmap(pte); 2508 __SetPageUptodate(new_page); 2509 pgtable = pmd_pgtable(_pmd); 2510 2511 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot); 2512 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma); 2513 2514 /* 2515 * spin_lock() below is not the equivalent of smp_wmb(), so 2516 * this is needed to avoid the copy_huge_page writes to become 2517 * visible after the set_pmd_at() write. 2518 */ 2519 smp_wmb(); 2520 2521 spin_lock(pmd_ptl); 2522 BUG_ON(!pmd_none(*pmd)); 2523 page_add_new_anon_rmap(new_page, vma, address); 2524 pgtable_trans_huge_deposit(mm, pmd, pgtable); 2525 set_pmd_at(mm, address, pmd, _pmd); 2526 update_mmu_cache_pmd(vma, address, pmd); 2527 spin_unlock(pmd_ptl); 2528 2529 *hpage = NULL; 2530 2531 khugepaged_pages_collapsed++; 2532 out_up_write: 2533 up_write(&mm->mmap_sem); 2534 return; 2535 2536 out: 2537 mem_cgroup_uncharge_page(new_page); 2538 goto out_up_write; 2539 } 2540 2541 static int khugepaged_scan_pmd(struct mm_struct *mm, 2542 struct vm_area_struct *vma, 2543 unsigned long address, 2544 struct page **hpage) 2545 { 2546 pmd_t *pmd; 2547 pte_t *pte, *_pte; 2548 int ret = 0, referenced = 0, none = 0; 2549 struct page *page; 2550 unsigned long _address; 2551 spinlock_t *ptl; 2552 int node = NUMA_NO_NODE; 2553 2554 VM_BUG_ON(address & ~HPAGE_PMD_MASK); 2555 2556 pmd = mm_find_pmd(mm, address); 2557 if (!pmd) 2558 goto out; 2559 if (pmd_trans_huge(*pmd)) 2560 goto out; 2561 2562 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load)); 2563 pte = pte_offset_map_lock(mm, pmd, address, &ptl); 2564 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR; 2565 _pte++, _address += PAGE_SIZE) { 2566 pte_t pteval = *_pte; 2567 if (pte_none(pteval)) { 2568 if (++none <= khugepaged_max_ptes_none) 2569 continue; 2570 else 2571 goto out_unmap; 2572 } 2573 if (!pte_present(pteval) || !pte_write(pteval)) 2574 goto out_unmap; 2575 page = vm_normal_page(vma, _address, pteval); 2576 if (unlikely(!page)) 2577 goto out_unmap; 2578 /* 2579 * Record which node the original page is from and save this 2580 * information to khugepaged_node_load[]. 2581 * Khupaged will allocate hugepage from the node has the max 2582 * hit record. 2583 */ 2584 node = page_to_nid(page); 2585 khugepaged_node_load[node]++; 2586 VM_BUG_ON(PageCompound(page)); 2587 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page)) 2588 goto out_unmap; 2589 /* cannot use mapcount: can't collapse if there's a gup pin */ 2590 if (page_count(page) != 1) 2591 goto out_unmap; 2592 if (pte_young(pteval) || PageReferenced(page) || 2593 mmu_notifier_test_young(vma->vm_mm, address)) 2594 referenced = 1; 2595 } 2596 if (referenced) 2597 ret = 1; 2598 out_unmap: 2599 pte_unmap_unlock(pte, ptl); 2600 if (ret) { 2601 node = khugepaged_find_target_node(); 2602 /* collapse_huge_page will return with the mmap_sem released */ 2603 collapse_huge_page(mm, address, hpage, vma, node); 2604 } 2605 out: 2606 return ret; 2607 } 2608 2609 static void collect_mm_slot(struct mm_slot *mm_slot) 2610 { 2611 struct mm_struct *mm = mm_slot->mm; 2612 2613 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock)); 2614 2615 if (khugepaged_test_exit(mm)) { 2616 /* free mm_slot */ 2617 hash_del(&mm_slot->hash); 2618 list_del(&mm_slot->mm_node); 2619 2620 /* 2621 * Not strictly needed because the mm exited already. 2622 * 2623 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags); 2624 */ 2625 2626 /* khugepaged_mm_lock actually not necessary for the below */ 2627 free_mm_slot(mm_slot); 2628 mmdrop(mm); 2629 } 2630 } 2631 2632 static unsigned int khugepaged_scan_mm_slot(unsigned int pages, 2633 struct page **hpage) 2634 __releases(&khugepaged_mm_lock) 2635 __acquires(&khugepaged_mm_lock) 2636 { 2637 struct mm_slot *mm_slot; 2638 struct mm_struct *mm; 2639 struct vm_area_struct *vma; 2640 int progress = 0; 2641 2642 VM_BUG_ON(!pages); 2643 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock)); 2644 2645 if (khugepaged_scan.mm_slot) 2646 mm_slot = khugepaged_scan.mm_slot; 2647 else { 2648 mm_slot = list_entry(khugepaged_scan.mm_head.next, 2649 struct mm_slot, mm_node); 2650 khugepaged_scan.address = 0; 2651 khugepaged_scan.mm_slot = mm_slot; 2652 } 2653 spin_unlock(&khugepaged_mm_lock); 2654 2655 mm = mm_slot->mm; 2656 down_read(&mm->mmap_sem); 2657 if (unlikely(khugepaged_test_exit(mm))) 2658 vma = NULL; 2659 else 2660 vma = find_vma(mm, khugepaged_scan.address); 2661 2662 progress++; 2663 for (; vma; vma = vma->vm_next) { 2664 unsigned long hstart, hend; 2665 2666 cond_resched(); 2667 if (unlikely(khugepaged_test_exit(mm))) { 2668 progress++; 2669 break; 2670 } 2671 if (!hugepage_vma_check(vma)) { 2672 skip: 2673 progress++; 2674 continue; 2675 } 2676 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; 2677 hend = vma->vm_end & HPAGE_PMD_MASK; 2678 if (hstart >= hend) 2679 goto skip; 2680 if (khugepaged_scan.address > hend) 2681 goto skip; 2682 if (khugepaged_scan.address < hstart) 2683 khugepaged_scan.address = hstart; 2684 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK); 2685 2686 while (khugepaged_scan.address < hend) { 2687 int ret; 2688 cond_resched(); 2689 if (unlikely(khugepaged_test_exit(mm))) 2690 goto breakouterloop; 2691 2692 VM_BUG_ON(khugepaged_scan.address < hstart || 2693 khugepaged_scan.address + HPAGE_PMD_SIZE > 2694 hend); 2695 ret = khugepaged_scan_pmd(mm, vma, 2696 khugepaged_scan.address, 2697 hpage); 2698 /* move to next address */ 2699 khugepaged_scan.address += HPAGE_PMD_SIZE; 2700 progress += HPAGE_PMD_NR; 2701 if (ret) 2702 /* we released mmap_sem so break loop */ 2703 goto breakouterloop_mmap_sem; 2704 if (progress >= pages) 2705 goto breakouterloop; 2706 } 2707 } 2708 breakouterloop: 2709 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */ 2710 breakouterloop_mmap_sem: 2711 2712 spin_lock(&khugepaged_mm_lock); 2713 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot); 2714 /* 2715 * Release the current mm_slot if this mm is about to die, or 2716 * if we scanned all vmas of this mm. 2717 */ 2718 if (khugepaged_test_exit(mm) || !vma) { 2719 /* 2720 * Make sure that if mm_users is reaching zero while 2721 * khugepaged runs here, khugepaged_exit will find 2722 * mm_slot not pointing to the exiting mm. 2723 */ 2724 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) { 2725 khugepaged_scan.mm_slot = list_entry( 2726 mm_slot->mm_node.next, 2727 struct mm_slot, mm_node); 2728 khugepaged_scan.address = 0; 2729 } else { 2730 khugepaged_scan.mm_slot = NULL; 2731 khugepaged_full_scans++; 2732 } 2733 2734 collect_mm_slot(mm_slot); 2735 } 2736 2737 return progress; 2738 } 2739 2740 static int khugepaged_has_work(void) 2741 { 2742 return !list_empty(&khugepaged_scan.mm_head) && 2743 khugepaged_enabled(); 2744 } 2745 2746 static int khugepaged_wait_event(void) 2747 { 2748 return !list_empty(&khugepaged_scan.mm_head) || 2749 kthread_should_stop(); 2750 } 2751 2752 static void khugepaged_do_scan(void) 2753 { 2754 struct page *hpage = NULL; 2755 unsigned int progress = 0, pass_through_head = 0; 2756 unsigned int pages = khugepaged_pages_to_scan; 2757 bool wait = true; 2758 2759 barrier(); /* write khugepaged_pages_to_scan to local stack */ 2760 2761 while (progress < pages) { 2762 if (!khugepaged_prealloc_page(&hpage, &wait)) 2763 break; 2764 2765 cond_resched(); 2766 2767 if (unlikely(kthread_should_stop() || freezing(current))) 2768 break; 2769 2770 spin_lock(&khugepaged_mm_lock); 2771 if (!khugepaged_scan.mm_slot) 2772 pass_through_head++; 2773 if (khugepaged_has_work() && 2774 pass_through_head < 2) 2775 progress += khugepaged_scan_mm_slot(pages - progress, 2776 &hpage); 2777 else 2778 progress = pages; 2779 spin_unlock(&khugepaged_mm_lock); 2780 } 2781 2782 if (!IS_ERR_OR_NULL(hpage)) 2783 put_page(hpage); 2784 } 2785 2786 static void khugepaged_wait_work(void) 2787 { 2788 try_to_freeze(); 2789 2790 if (khugepaged_has_work()) { 2791 if (!khugepaged_scan_sleep_millisecs) 2792 return; 2793 2794 wait_event_freezable_timeout(khugepaged_wait, 2795 kthread_should_stop(), 2796 msecs_to_jiffies(khugepaged_scan_sleep_millisecs)); 2797 return; 2798 } 2799 2800 if (khugepaged_enabled()) 2801 wait_event_freezable(khugepaged_wait, khugepaged_wait_event()); 2802 } 2803 2804 static int khugepaged(void *none) 2805 { 2806 struct mm_slot *mm_slot; 2807 2808 set_freezable(); 2809 set_user_nice(current, 19); 2810 2811 while (!kthread_should_stop()) { 2812 khugepaged_do_scan(); 2813 khugepaged_wait_work(); 2814 } 2815 2816 spin_lock(&khugepaged_mm_lock); 2817 mm_slot = khugepaged_scan.mm_slot; 2818 khugepaged_scan.mm_slot = NULL; 2819 if (mm_slot) 2820 collect_mm_slot(mm_slot); 2821 spin_unlock(&khugepaged_mm_lock); 2822 return 0; 2823 } 2824 2825 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, 2826 unsigned long haddr, pmd_t *pmd) 2827 { 2828 struct mm_struct *mm = vma->vm_mm; 2829 pgtable_t pgtable; 2830 pmd_t _pmd; 2831 int i; 2832 2833 pmdp_clear_flush(vma, haddr, pmd); 2834 /* leave pmd empty until pte is filled */ 2835 2836 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 2837 pmd_populate(mm, &_pmd, pgtable); 2838 2839 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 2840 pte_t *pte, entry; 2841 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot); 2842 entry = pte_mkspecial(entry); 2843 pte = pte_offset_map(&_pmd, haddr); 2844 VM_BUG_ON(!pte_none(*pte)); 2845 set_pte_at(mm, haddr, pte, entry); 2846 pte_unmap(pte); 2847 } 2848 smp_wmb(); /* make pte visible before pmd */ 2849 pmd_populate(mm, pmd, pgtable); 2850 put_huge_zero_page(); 2851 } 2852 2853 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address, 2854 pmd_t *pmd) 2855 { 2856 spinlock_t *ptl; 2857 struct page *page; 2858 struct mm_struct *mm = vma->vm_mm; 2859 unsigned long haddr = address & HPAGE_PMD_MASK; 2860 unsigned long mmun_start; /* For mmu_notifiers */ 2861 unsigned long mmun_end; /* For mmu_notifiers */ 2862 2863 BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE); 2864 2865 mmun_start = haddr; 2866 mmun_end = haddr + HPAGE_PMD_SIZE; 2867 again: 2868 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 2869 ptl = pmd_lock(mm, pmd); 2870 if (unlikely(!pmd_trans_huge(*pmd))) { 2871 spin_unlock(ptl); 2872 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 2873 return; 2874 } 2875 if (is_huge_zero_pmd(*pmd)) { 2876 __split_huge_zero_page_pmd(vma, haddr, pmd); 2877 spin_unlock(ptl); 2878 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 2879 return; 2880 } 2881 page = pmd_page(*pmd); 2882 VM_BUG_ON(!page_count(page)); 2883 get_page(page); 2884 spin_unlock(ptl); 2885 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 2886 2887 split_huge_page(page); 2888 2889 put_page(page); 2890 2891 /* 2892 * We don't always have down_write of mmap_sem here: a racing 2893 * do_huge_pmd_wp_page() might have copied-on-write to another 2894 * huge page before our split_huge_page() got the anon_vma lock. 2895 */ 2896 if (unlikely(pmd_trans_huge(*pmd))) 2897 goto again; 2898 } 2899 2900 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address, 2901 pmd_t *pmd) 2902 { 2903 struct vm_area_struct *vma; 2904 2905 vma = find_vma(mm, address); 2906 BUG_ON(vma == NULL); 2907 split_huge_page_pmd(vma, address, pmd); 2908 } 2909 2910 static void split_huge_page_address(struct mm_struct *mm, 2911 unsigned long address) 2912 { 2913 pmd_t *pmd; 2914 2915 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK)); 2916 2917 pmd = mm_find_pmd(mm, address); 2918 if (!pmd) 2919 return; 2920 /* 2921 * Caller holds the mmap_sem write mode, so a huge pmd cannot 2922 * materialize from under us. 2923 */ 2924 split_huge_page_pmd_mm(mm, address, pmd); 2925 } 2926 2927 void __vma_adjust_trans_huge(struct vm_area_struct *vma, 2928 unsigned long start, 2929 unsigned long end, 2930 long adjust_next) 2931 { 2932 /* 2933 * If the new start address isn't hpage aligned and it could 2934 * previously contain an hugepage: check if we need to split 2935 * an huge pmd. 2936 */ 2937 if (start & ~HPAGE_PMD_MASK && 2938 (start & HPAGE_PMD_MASK) >= vma->vm_start && 2939 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) 2940 split_huge_page_address(vma->vm_mm, start); 2941 2942 /* 2943 * If the new end address isn't hpage aligned and it could 2944 * previously contain an hugepage: check if we need to split 2945 * an huge pmd. 2946 */ 2947 if (end & ~HPAGE_PMD_MASK && 2948 (end & HPAGE_PMD_MASK) >= vma->vm_start && 2949 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) 2950 split_huge_page_address(vma->vm_mm, end); 2951 2952 /* 2953 * If we're also updating the vma->vm_next->vm_start, if the new 2954 * vm_next->vm_start isn't page aligned and it could previously 2955 * contain an hugepage: check if we need to split an huge pmd. 2956 */ 2957 if (adjust_next > 0) { 2958 struct vm_area_struct *next = vma->vm_next; 2959 unsigned long nstart = next->vm_start; 2960 nstart += adjust_next << PAGE_SHIFT; 2961 if (nstart & ~HPAGE_PMD_MASK && 2962 (nstart & HPAGE_PMD_MASK) >= next->vm_start && 2963 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end) 2964 split_huge_page_address(next->vm_mm, nstart); 2965 } 2966 } 2967