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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 9 10 #include <linux/mm.h> 11 #include <linux/sched.h> 12 #include <linux/highmem.h> 13 #include <linux/hugetlb.h> 14 #include <linux/mmu_notifier.h> 15 #include <linux/rmap.h> 16 #include <linux/swap.h> 17 #include <linux/shrinker.h> 18 #include <linux/mm_inline.h> 19 #include <linux/dax.h> 20 #include <linux/kthread.h> 21 #include <linux/khugepaged.h> 22 #include <linux/freezer.h> 23 #include <linux/mman.h> 24 #include <linux/pagemap.h> 25 #include <linux/migrate.h> 26 #include <linux/hashtable.h> 27 #include <linux/userfaultfd_k.h> 28 #include <linux/page_idle.h> 29 30 #include <asm/tlb.h> 31 #include <asm/pgalloc.h> 32 #include "internal.h" 33 34 /* 35 * By default transparent hugepage support is disabled in order that avoid 36 * to risk increase the memory footprint of applications without a guaranteed 37 * benefit. When transparent hugepage support is enabled, is for all mappings, 38 * and khugepaged scans all mappings. 39 * Defrag is invoked by khugepaged hugepage allocations and by page faults 40 * for all hugepage allocations. 41 */ 42 unsigned long transparent_hugepage_flags __read_mostly = 43 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS 44 (1<<TRANSPARENT_HUGEPAGE_FLAG)| 45 #endif 46 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE 47 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)| 48 #endif 49 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)| 50 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)| 51 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 52 53 /* default scan 8*512 pte (or vmas) every 30 second */ 54 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8; 55 static unsigned int khugepaged_pages_collapsed; 56 static unsigned int khugepaged_full_scans; 57 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000; 58 /* during fragmentation poll the hugepage allocator once every minute */ 59 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000; 60 static struct task_struct *khugepaged_thread __read_mostly; 61 static DEFINE_MUTEX(khugepaged_mutex); 62 static DEFINE_SPINLOCK(khugepaged_mm_lock); 63 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait); 64 /* 65 * default collapse hugepages if there is at least one pte mapped like 66 * it would have happened if the vma was large enough during page 67 * fault. 68 */ 69 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1; 70 71 static int khugepaged(void *none); 72 static int khugepaged_slab_init(void); 73 static void khugepaged_slab_exit(void); 74 75 #define MM_SLOTS_HASH_BITS 10 76 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); 77 78 static struct kmem_cache *mm_slot_cache __read_mostly; 79 80 /** 81 * struct mm_slot - hash lookup from mm to mm_slot 82 * @hash: hash collision list 83 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head 84 * @mm: the mm that this information is valid for 85 */ 86 struct mm_slot { 87 struct hlist_node hash; 88 struct list_head mm_node; 89 struct mm_struct *mm; 90 }; 91 92 /** 93 * struct khugepaged_scan - cursor for scanning 94 * @mm_head: the head of the mm list to scan 95 * @mm_slot: the current mm_slot we are scanning 96 * @address: the next address inside that to be scanned 97 * 98 * There is only the one khugepaged_scan instance of this cursor structure. 99 */ 100 struct khugepaged_scan { 101 struct list_head mm_head; 102 struct mm_slot *mm_slot; 103 unsigned long address; 104 }; 105 static struct khugepaged_scan khugepaged_scan = { 106 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head), 107 }; 108 109 110 static void set_recommended_min_free_kbytes(void) 111 { 112 struct zone *zone; 113 int nr_zones = 0; 114 unsigned long recommended_min; 115 116 for_each_populated_zone(zone) 117 nr_zones++; 118 119 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */ 120 recommended_min = pageblock_nr_pages * nr_zones * 2; 121 122 /* 123 * Make sure that on average at least two pageblocks are almost free 124 * of another type, one for a migratetype to fall back to and a 125 * second to avoid subsequent fallbacks of other types There are 3 126 * MIGRATE_TYPES we care about. 127 */ 128 recommended_min += pageblock_nr_pages * nr_zones * 129 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES; 130 131 /* don't ever allow to reserve more than 5% of the lowmem */ 132 recommended_min = min(recommended_min, 133 (unsigned long) nr_free_buffer_pages() / 20); 134 recommended_min <<= (PAGE_SHIFT-10); 135 136 if (recommended_min > min_free_kbytes) { 137 if (user_min_free_kbytes >= 0) 138 pr_info("raising min_free_kbytes from %d to %lu " 139 "to help transparent hugepage allocations\n", 140 min_free_kbytes, recommended_min); 141 142 min_free_kbytes = recommended_min; 143 } 144 setup_per_zone_wmarks(); 145 } 146 147 static int start_stop_khugepaged(void) 148 { 149 int err = 0; 150 if (khugepaged_enabled()) { 151 if (!khugepaged_thread) 152 khugepaged_thread = kthread_run(khugepaged, NULL, 153 "khugepaged"); 154 if (IS_ERR(khugepaged_thread)) { 155 pr_err("khugepaged: kthread_run(khugepaged) failed\n"); 156 err = PTR_ERR(khugepaged_thread); 157 khugepaged_thread = NULL; 158 goto fail; 159 } 160 161 if (!list_empty(&khugepaged_scan.mm_head)) 162 wake_up_interruptible(&khugepaged_wait); 163 164 set_recommended_min_free_kbytes(); 165 } else if (khugepaged_thread) { 166 kthread_stop(khugepaged_thread); 167 khugepaged_thread = NULL; 168 } 169 fail: 170 return err; 171 } 172 173 static atomic_t huge_zero_refcount; 174 struct page *huge_zero_page __read_mostly; 175 176 struct page *get_huge_zero_page(void) 177 { 178 struct page *zero_page; 179 retry: 180 if (likely(atomic_inc_not_zero(&huge_zero_refcount))) 181 return READ_ONCE(huge_zero_page); 182 183 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE, 184 HPAGE_PMD_ORDER); 185 if (!zero_page) { 186 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED); 187 return NULL; 188 } 189 count_vm_event(THP_ZERO_PAGE_ALLOC); 190 preempt_disable(); 191 if (cmpxchg(&huge_zero_page, NULL, zero_page)) { 192 preempt_enable(); 193 __free_pages(zero_page, compound_order(zero_page)); 194 goto retry; 195 } 196 197 /* We take additional reference here. It will be put back by shrinker */ 198 atomic_set(&huge_zero_refcount, 2); 199 preempt_enable(); 200 return READ_ONCE(huge_zero_page); 201 } 202 203 static void put_huge_zero_page(void) 204 { 205 /* 206 * Counter should never go to zero here. Only shrinker can put 207 * last reference. 208 */ 209 BUG_ON(atomic_dec_and_test(&huge_zero_refcount)); 210 } 211 212 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink, 213 struct shrink_control *sc) 214 { 215 /* we can free zero page only if last reference remains */ 216 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0; 217 } 218 219 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink, 220 struct shrink_control *sc) 221 { 222 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) { 223 struct page *zero_page = xchg(&huge_zero_page, NULL); 224 BUG_ON(zero_page == NULL); 225 __free_pages(zero_page, compound_order(zero_page)); 226 return HPAGE_PMD_NR; 227 } 228 229 return 0; 230 } 231 232 static struct shrinker huge_zero_page_shrinker = { 233 .count_objects = shrink_huge_zero_page_count, 234 .scan_objects = shrink_huge_zero_page_scan, 235 .seeks = DEFAULT_SEEKS, 236 }; 237 238 #ifdef CONFIG_SYSFS 239 240 static ssize_t double_flag_show(struct kobject *kobj, 241 struct kobj_attribute *attr, char *buf, 242 enum transparent_hugepage_flag enabled, 243 enum transparent_hugepage_flag req_madv) 244 { 245 if (test_bit(enabled, &transparent_hugepage_flags)) { 246 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags)); 247 return sprintf(buf, "[always] madvise never\n"); 248 } else if (test_bit(req_madv, &transparent_hugepage_flags)) 249 return sprintf(buf, "always [madvise] never\n"); 250 else 251 return sprintf(buf, "always madvise [never]\n"); 252 } 253 static ssize_t double_flag_store(struct kobject *kobj, 254 struct kobj_attribute *attr, 255 const char *buf, size_t count, 256 enum transparent_hugepage_flag enabled, 257 enum transparent_hugepage_flag req_madv) 258 { 259 if (!memcmp("always", buf, 260 min(sizeof("always")-1, count))) { 261 set_bit(enabled, &transparent_hugepage_flags); 262 clear_bit(req_madv, &transparent_hugepage_flags); 263 } else if (!memcmp("madvise", buf, 264 min(sizeof("madvise")-1, count))) { 265 clear_bit(enabled, &transparent_hugepage_flags); 266 set_bit(req_madv, &transparent_hugepage_flags); 267 } else if (!memcmp("never", buf, 268 min(sizeof("never")-1, count))) { 269 clear_bit(enabled, &transparent_hugepage_flags); 270 clear_bit(req_madv, &transparent_hugepage_flags); 271 } else 272 return -EINVAL; 273 274 return count; 275 } 276 277 static ssize_t enabled_show(struct kobject *kobj, 278 struct kobj_attribute *attr, char *buf) 279 { 280 return double_flag_show(kobj, attr, buf, 281 TRANSPARENT_HUGEPAGE_FLAG, 282 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG); 283 } 284 static ssize_t enabled_store(struct kobject *kobj, 285 struct kobj_attribute *attr, 286 const char *buf, size_t count) 287 { 288 ssize_t ret; 289 290 ret = double_flag_store(kobj, attr, buf, count, 291 TRANSPARENT_HUGEPAGE_FLAG, 292 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG); 293 294 if (ret > 0) { 295 int err; 296 297 mutex_lock(&khugepaged_mutex); 298 err = start_stop_khugepaged(); 299 mutex_unlock(&khugepaged_mutex); 300 301 if (err) 302 ret = err; 303 } 304 305 return ret; 306 } 307 static struct kobj_attribute enabled_attr = 308 __ATTR(enabled, 0644, enabled_show, enabled_store); 309 310 static ssize_t single_flag_show(struct kobject *kobj, 311 struct kobj_attribute *attr, char *buf, 312 enum transparent_hugepage_flag flag) 313 { 314 return sprintf(buf, "%d\n", 315 !!test_bit(flag, &transparent_hugepage_flags)); 316 } 317 318 static ssize_t single_flag_store(struct kobject *kobj, 319 struct kobj_attribute *attr, 320 const char *buf, size_t count, 321 enum transparent_hugepage_flag flag) 322 { 323 unsigned long value; 324 int ret; 325 326 ret = kstrtoul(buf, 10, &value); 327 if (ret < 0) 328 return ret; 329 if (value > 1) 330 return -EINVAL; 331 332 if (value) 333 set_bit(flag, &transparent_hugepage_flags); 334 else 335 clear_bit(flag, &transparent_hugepage_flags); 336 337 return count; 338 } 339 340 /* 341 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind 342 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of 343 * memory just to allocate one more hugepage. 344 */ 345 static ssize_t defrag_show(struct kobject *kobj, 346 struct kobj_attribute *attr, char *buf) 347 { 348 return double_flag_show(kobj, attr, buf, 349 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG, 350 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG); 351 } 352 static ssize_t defrag_store(struct kobject *kobj, 353 struct kobj_attribute *attr, 354 const char *buf, size_t count) 355 { 356 return double_flag_store(kobj, attr, buf, count, 357 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG, 358 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG); 359 } 360 static struct kobj_attribute defrag_attr = 361 __ATTR(defrag, 0644, defrag_show, defrag_store); 362 363 static ssize_t use_zero_page_show(struct kobject *kobj, 364 struct kobj_attribute *attr, char *buf) 365 { 366 return single_flag_show(kobj, attr, buf, 367 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 368 } 369 static ssize_t use_zero_page_store(struct kobject *kobj, 370 struct kobj_attribute *attr, const char *buf, size_t count) 371 { 372 return single_flag_store(kobj, attr, buf, count, 373 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 374 } 375 static struct kobj_attribute use_zero_page_attr = 376 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store); 377 #ifdef CONFIG_DEBUG_VM 378 static ssize_t debug_cow_show(struct kobject *kobj, 379 struct kobj_attribute *attr, char *buf) 380 { 381 return single_flag_show(kobj, attr, buf, 382 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); 383 } 384 static ssize_t debug_cow_store(struct kobject *kobj, 385 struct kobj_attribute *attr, 386 const char *buf, size_t count) 387 { 388 return single_flag_store(kobj, attr, buf, count, 389 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); 390 } 391 static struct kobj_attribute debug_cow_attr = 392 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store); 393 #endif /* CONFIG_DEBUG_VM */ 394 395 static struct attribute *hugepage_attr[] = { 396 &enabled_attr.attr, 397 &defrag_attr.attr, 398 &use_zero_page_attr.attr, 399 #ifdef CONFIG_DEBUG_VM 400 &debug_cow_attr.attr, 401 #endif 402 NULL, 403 }; 404 405 static struct attribute_group hugepage_attr_group = { 406 .attrs = hugepage_attr, 407 }; 408 409 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj, 410 struct kobj_attribute *attr, 411 char *buf) 412 { 413 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs); 414 } 415 416 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj, 417 struct kobj_attribute *attr, 418 const char *buf, size_t count) 419 { 420 unsigned long msecs; 421 int err; 422 423 err = kstrtoul(buf, 10, &msecs); 424 if (err || msecs > UINT_MAX) 425 return -EINVAL; 426 427 khugepaged_scan_sleep_millisecs = msecs; 428 wake_up_interruptible(&khugepaged_wait); 429 430 return count; 431 } 432 static struct kobj_attribute scan_sleep_millisecs_attr = 433 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show, 434 scan_sleep_millisecs_store); 435 436 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj, 437 struct kobj_attribute *attr, 438 char *buf) 439 { 440 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs); 441 } 442 443 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj, 444 struct kobj_attribute *attr, 445 const char *buf, size_t count) 446 { 447 unsigned long msecs; 448 int err; 449 450 err = kstrtoul(buf, 10, &msecs); 451 if (err || msecs > UINT_MAX) 452 return -EINVAL; 453 454 khugepaged_alloc_sleep_millisecs = msecs; 455 wake_up_interruptible(&khugepaged_wait); 456 457 return count; 458 } 459 static struct kobj_attribute alloc_sleep_millisecs_attr = 460 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show, 461 alloc_sleep_millisecs_store); 462 463 static ssize_t pages_to_scan_show(struct kobject *kobj, 464 struct kobj_attribute *attr, 465 char *buf) 466 { 467 return sprintf(buf, "%u\n", khugepaged_pages_to_scan); 468 } 469 static ssize_t pages_to_scan_store(struct kobject *kobj, 470 struct kobj_attribute *attr, 471 const char *buf, size_t count) 472 { 473 int err; 474 unsigned long pages; 475 476 err = kstrtoul(buf, 10, &pages); 477 if (err || !pages || pages > UINT_MAX) 478 return -EINVAL; 479 480 khugepaged_pages_to_scan = pages; 481 482 return count; 483 } 484 static struct kobj_attribute pages_to_scan_attr = 485 __ATTR(pages_to_scan, 0644, pages_to_scan_show, 486 pages_to_scan_store); 487 488 static ssize_t pages_collapsed_show(struct kobject *kobj, 489 struct kobj_attribute *attr, 490 char *buf) 491 { 492 return sprintf(buf, "%u\n", khugepaged_pages_collapsed); 493 } 494 static struct kobj_attribute pages_collapsed_attr = 495 __ATTR_RO(pages_collapsed); 496 497 static ssize_t full_scans_show(struct kobject *kobj, 498 struct kobj_attribute *attr, 499 char *buf) 500 { 501 return sprintf(buf, "%u\n", khugepaged_full_scans); 502 } 503 static struct kobj_attribute full_scans_attr = 504 __ATTR_RO(full_scans); 505 506 static ssize_t khugepaged_defrag_show(struct kobject *kobj, 507 struct kobj_attribute *attr, char *buf) 508 { 509 return single_flag_show(kobj, attr, buf, 510 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); 511 } 512 static ssize_t khugepaged_defrag_store(struct kobject *kobj, 513 struct kobj_attribute *attr, 514 const char *buf, size_t count) 515 { 516 return single_flag_store(kobj, attr, buf, count, 517 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); 518 } 519 static struct kobj_attribute khugepaged_defrag_attr = 520 __ATTR(defrag, 0644, khugepaged_defrag_show, 521 khugepaged_defrag_store); 522 523 /* 524 * max_ptes_none controls if khugepaged should collapse hugepages over 525 * any unmapped ptes in turn potentially increasing the memory 526 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not 527 * reduce the available free memory in the system as it 528 * runs. Increasing max_ptes_none will instead potentially reduce the 529 * free memory in the system during the khugepaged scan. 530 */ 531 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj, 532 struct kobj_attribute *attr, 533 char *buf) 534 { 535 return sprintf(buf, "%u\n", khugepaged_max_ptes_none); 536 } 537 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj, 538 struct kobj_attribute *attr, 539 const char *buf, size_t count) 540 { 541 int err; 542 unsigned long max_ptes_none; 543 544 err = kstrtoul(buf, 10, &max_ptes_none); 545 if (err || max_ptes_none > HPAGE_PMD_NR-1) 546 return -EINVAL; 547 548 khugepaged_max_ptes_none = max_ptes_none; 549 550 return count; 551 } 552 static struct kobj_attribute khugepaged_max_ptes_none_attr = 553 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show, 554 khugepaged_max_ptes_none_store); 555 556 static struct attribute *khugepaged_attr[] = { 557 &khugepaged_defrag_attr.attr, 558 &khugepaged_max_ptes_none_attr.attr, 559 &pages_to_scan_attr.attr, 560 &pages_collapsed_attr.attr, 561 &full_scans_attr.attr, 562 &scan_sleep_millisecs_attr.attr, 563 &alloc_sleep_millisecs_attr.attr, 564 NULL, 565 }; 566 567 static struct attribute_group khugepaged_attr_group = { 568 .attrs = khugepaged_attr, 569 .name = "khugepaged", 570 }; 571 572 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj) 573 { 574 int err; 575 576 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj); 577 if (unlikely(!*hugepage_kobj)) { 578 pr_err("failed to create transparent hugepage kobject\n"); 579 return -ENOMEM; 580 } 581 582 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group); 583 if (err) { 584 pr_err("failed to register transparent hugepage group\n"); 585 goto delete_obj; 586 } 587 588 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group); 589 if (err) { 590 pr_err("failed to register transparent hugepage group\n"); 591 goto remove_hp_group; 592 } 593 594 return 0; 595 596 remove_hp_group: 597 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group); 598 delete_obj: 599 kobject_put(*hugepage_kobj); 600 return err; 601 } 602 603 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj) 604 { 605 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group); 606 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group); 607 kobject_put(hugepage_kobj); 608 } 609 #else 610 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj) 611 { 612 return 0; 613 } 614 615 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj) 616 { 617 } 618 #endif /* CONFIG_SYSFS */ 619 620 static int __init hugepage_init(void) 621 { 622 int err; 623 struct kobject *hugepage_kobj; 624 625 if (!has_transparent_hugepage()) { 626 transparent_hugepage_flags = 0; 627 return -EINVAL; 628 } 629 630 err = hugepage_init_sysfs(&hugepage_kobj); 631 if (err) 632 goto err_sysfs; 633 634 err = khugepaged_slab_init(); 635 if (err) 636 goto err_slab; 637 638 err = register_shrinker(&huge_zero_page_shrinker); 639 if (err) 640 goto err_hzp_shrinker; 641 642 /* 643 * By default disable transparent hugepages on smaller systems, 644 * where the extra memory used could hurt more than TLB overhead 645 * is likely to save. The admin can still enable it through /sys. 646 */ 647 if (totalram_pages < (512 << (20 - PAGE_SHIFT))) { 648 transparent_hugepage_flags = 0; 649 return 0; 650 } 651 652 err = start_stop_khugepaged(); 653 if (err) 654 goto err_khugepaged; 655 656 return 0; 657 err_khugepaged: 658 unregister_shrinker(&huge_zero_page_shrinker); 659 err_hzp_shrinker: 660 khugepaged_slab_exit(); 661 err_slab: 662 hugepage_exit_sysfs(hugepage_kobj); 663 err_sysfs: 664 return err; 665 } 666 subsys_initcall(hugepage_init); 667 668 static int __init setup_transparent_hugepage(char *str) 669 { 670 int ret = 0; 671 if (!str) 672 goto out; 673 if (!strcmp(str, "always")) { 674 set_bit(TRANSPARENT_HUGEPAGE_FLAG, 675 &transparent_hugepage_flags); 676 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 677 &transparent_hugepage_flags); 678 ret = 1; 679 } else if (!strcmp(str, "madvise")) { 680 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 681 &transparent_hugepage_flags); 682 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 683 &transparent_hugepage_flags); 684 ret = 1; 685 } else if (!strcmp(str, "never")) { 686 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 687 &transparent_hugepage_flags); 688 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 689 &transparent_hugepage_flags); 690 ret = 1; 691 } 692 out: 693 if (!ret) 694 pr_warn("transparent_hugepage= cannot parse, ignored\n"); 695 return ret; 696 } 697 __setup("transparent_hugepage=", setup_transparent_hugepage); 698 699 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma) 700 { 701 if (likely(vma->vm_flags & VM_WRITE)) 702 pmd = pmd_mkwrite(pmd); 703 return pmd; 704 } 705 706 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot) 707 { 708 pmd_t entry; 709 entry = mk_pmd(page, prot); 710 entry = pmd_mkhuge(entry); 711 return entry; 712 } 713 714 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm, 715 struct vm_area_struct *vma, 716 unsigned long address, pmd_t *pmd, 717 struct page *page, gfp_t gfp, 718 unsigned int flags) 719 { 720 struct mem_cgroup *memcg; 721 pgtable_t pgtable; 722 spinlock_t *ptl; 723 unsigned long haddr = address & HPAGE_PMD_MASK; 724 725 VM_BUG_ON_PAGE(!PageCompound(page), page); 726 727 if (mem_cgroup_try_charge(page, mm, gfp, &memcg)) { 728 put_page(page); 729 count_vm_event(THP_FAULT_FALLBACK); 730 return VM_FAULT_FALLBACK; 731 } 732 733 pgtable = pte_alloc_one(mm, haddr); 734 if (unlikely(!pgtable)) { 735 mem_cgroup_cancel_charge(page, memcg); 736 put_page(page); 737 return VM_FAULT_OOM; 738 } 739 740 clear_huge_page(page, haddr, HPAGE_PMD_NR); 741 /* 742 * The memory barrier inside __SetPageUptodate makes sure that 743 * clear_huge_page writes become visible before the set_pmd_at() 744 * write. 745 */ 746 __SetPageUptodate(page); 747 748 ptl = pmd_lock(mm, pmd); 749 if (unlikely(!pmd_none(*pmd))) { 750 spin_unlock(ptl); 751 mem_cgroup_cancel_charge(page, memcg); 752 put_page(page); 753 pte_free(mm, pgtable); 754 } else { 755 pmd_t entry; 756 757 /* Deliver the page fault to userland */ 758 if (userfaultfd_missing(vma)) { 759 int ret; 760 761 spin_unlock(ptl); 762 mem_cgroup_cancel_charge(page, memcg); 763 put_page(page); 764 pte_free(mm, pgtable); 765 ret = handle_userfault(vma, address, flags, 766 VM_UFFD_MISSING); 767 VM_BUG_ON(ret & VM_FAULT_FALLBACK); 768 return ret; 769 } 770 771 entry = mk_huge_pmd(page, vma->vm_page_prot); 772 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 773 page_add_new_anon_rmap(page, vma, haddr); 774 mem_cgroup_commit_charge(page, memcg, false); 775 lru_cache_add_active_or_unevictable(page, vma); 776 pgtable_trans_huge_deposit(mm, pmd, pgtable); 777 set_pmd_at(mm, haddr, pmd, entry); 778 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR); 779 atomic_long_inc(&mm->nr_ptes); 780 spin_unlock(ptl); 781 count_vm_event(THP_FAULT_ALLOC); 782 } 783 784 return 0; 785 } 786 787 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp) 788 { 789 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_RECLAIM)) | extra_gfp; 790 } 791 792 /* Caller must hold page table lock. */ 793 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, 794 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, 795 struct page *zero_page) 796 { 797 pmd_t entry; 798 if (!pmd_none(*pmd)) 799 return false; 800 entry = mk_pmd(zero_page, vma->vm_page_prot); 801 entry = pmd_mkhuge(entry); 802 pgtable_trans_huge_deposit(mm, pmd, pgtable); 803 set_pmd_at(mm, haddr, pmd, entry); 804 atomic_long_inc(&mm->nr_ptes); 805 return true; 806 } 807 808 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma, 809 unsigned long address, pmd_t *pmd, 810 unsigned int flags) 811 { 812 gfp_t gfp; 813 struct page *page; 814 unsigned long haddr = address & HPAGE_PMD_MASK; 815 816 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end) 817 return VM_FAULT_FALLBACK; 818 if (unlikely(anon_vma_prepare(vma))) 819 return VM_FAULT_OOM; 820 if (unlikely(khugepaged_enter(vma, vma->vm_flags))) 821 return VM_FAULT_OOM; 822 if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) && 823 transparent_hugepage_use_zero_page()) { 824 spinlock_t *ptl; 825 pgtable_t pgtable; 826 struct page *zero_page; 827 bool set; 828 int ret; 829 pgtable = pte_alloc_one(mm, haddr); 830 if (unlikely(!pgtable)) 831 return VM_FAULT_OOM; 832 zero_page = get_huge_zero_page(); 833 if (unlikely(!zero_page)) { 834 pte_free(mm, pgtable); 835 count_vm_event(THP_FAULT_FALLBACK); 836 return VM_FAULT_FALLBACK; 837 } 838 ptl = pmd_lock(mm, pmd); 839 ret = 0; 840 set = false; 841 if (pmd_none(*pmd)) { 842 if (userfaultfd_missing(vma)) { 843 spin_unlock(ptl); 844 ret = handle_userfault(vma, address, flags, 845 VM_UFFD_MISSING); 846 VM_BUG_ON(ret & VM_FAULT_FALLBACK); 847 } else { 848 set_huge_zero_page(pgtable, mm, vma, 849 haddr, pmd, 850 zero_page); 851 spin_unlock(ptl); 852 set = true; 853 } 854 } else 855 spin_unlock(ptl); 856 if (!set) { 857 pte_free(mm, pgtable); 858 put_huge_zero_page(); 859 } 860 return ret; 861 } 862 gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0); 863 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER); 864 if (unlikely(!page)) { 865 count_vm_event(THP_FAULT_FALLBACK); 866 return VM_FAULT_FALLBACK; 867 } 868 return __do_huge_pmd_anonymous_page(mm, vma, address, pmd, page, gfp, 869 flags); 870 } 871 872 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, 873 pmd_t *pmd, unsigned long pfn, pgprot_t prot, bool write) 874 { 875 struct mm_struct *mm = vma->vm_mm; 876 pmd_t entry; 877 spinlock_t *ptl; 878 879 ptl = pmd_lock(mm, pmd); 880 if (pmd_none(*pmd)) { 881 entry = pmd_mkhuge(pfn_pmd(pfn, prot)); 882 if (write) { 883 entry = pmd_mkyoung(pmd_mkdirty(entry)); 884 entry = maybe_pmd_mkwrite(entry, vma); 885 } 886 set_pmd_at(mm, addr, pmd, entry); 887 update_mmu_cache_pmd(vma, addr, pmd); 888 } 889 spin_unlock(ptl); 890 } 891 892 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, 893 pmd_t *pmd, unsigned long pfn, bool write) 894 { 895 pgprot_t pgprot = vma->vm_page_prot; 896 /* 897 * If we had pmd_special, we could avoid all these restrictions, 898 * but we need to be consistent with PTEs and architectures that 899 * can't support a 'special' bit. 900 */ 901 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); 902 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 903 (VM_PFNMAP|VM_MIXEDMAP)); 904 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 905 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn)); 906 907 if (addr < vma->vm_start || addr >= vma->vm_end) 908 return VM_FAULT_SIGBUS; 909 if (track_pfn_insert(vma, &pgprot, pfn)) 910 return VM_FAULT_SIGBUS; 911 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write); 912 return VM_FAULT_NOPAGE; 913 } 914 915 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, 916 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, 917 struct vm_area_struct *vma) 918 { 919 spinlock_t *dst_ptl, *src_ptl; 920 struct page *src_page; 921 pmd_t pmd; 922 pgtable_t pgtable; 923 int ret; 924 925 ret = -ENOMEM; 926 pgtable = pte_alloc_one(dst_mm, addr); 927 if (unlikely(!pgtable)) 928 goto out; 929 930 dst_ptl = pmd_lock(dst_mm, dst_pmd); 931 src_ptl = pmd_lockptr(src_mm, src_pmd); 932 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 933 934 ret = -EAGAIN; 935 pmd = *src_pmd; 936 if (unlikely(!pmd_trans_huge(pmd))) { 937 pte_free(dst_mm, pgtable); 938 goto out_unlock; 939 } 940 /* 941 * When page table lock is held, the huge zero pmd should not be 942 * under splitting since we don't split the page itself, only pmd to 943 * a page table. 944 */ 945 if (is_huge_zero_pmd(pmd)) { 946 struct page *zero_page; 947 /* 948 * get_huge_zero_page() will never allocate a new page here, 949 * since we already have a zero page to copy. It just takes a 950 * reference. 951 */ 952 zero_page = get_huge_zero_page(); 953 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd, 954 zero_page); 955 ret = 0; 956 goto out_unlock; 957 } 958 959 if (unlikely(pmd_trans_splitting(pmd))) { 960 /* split huge page running from under us */ 961 spin_unlock(src_ptl); 962 spin_unlock(dst_ptl); 963 pte_free(dst_mm, pgtable); 964 965 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */ 966 goto out; 967 } 968 src_page = pmd_page(pmd); 969 VM_BUG_ON_PAGE(!PageHead(src_page), src_page); 970 get_page(src_page); 971 page_dup_rmap(src_page); 972 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 973 974 pmdp_set_wrprotect(src_mm, addr, src_pmd); 975 pmd = pmd_mkold(pmd_wrprotect(pmd)); 976 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); 977 set_pmd_at(dst_mm, addr, dst_pmd, pmd); 978 atomic_long_inc(&dst_mm->nr_ptes); 979 980 ret = 0; 981 out_unlock: 982 spin_unlock(src_ptl); 983 spin_unlock(dst_ptl); 984 out: 985 return ret; 986 } 987 988 void huge_pmd_set_accessed(struct mm_struct *mm, 989 struct vm_area_struct *vma, 990 unsigned long address, 991 pmd_t *pmd, pmd_t orig_pmd, 992 int dirty) 993 { 994 spinlock_t *ptl; 995 pmd_t entry; 996 unsigned long haddr; 997 998 ptl = pmd_lock(mm, pmd); 999 if (unlikely(!pmd_same(*pmd, orig_pmd))) 1000 goto unlock; 1001 1002 entry = pmd_mkyoung(orig_pmd); 1003 haddr = address & HPAGE_PMD_MASK; 1004 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty)) 1005 update_mmu_cache_pmd(vma, address, pmd); 1006 1007 unlock: 1008 spin_unlock(ptl); 1009 } 1010 1011 /* 1012 * Save CONFIG_DEBUG_PAGEALLOC from faulting falsely on tail pages 1013 * during copy_user_huge_page()'s copy_page_rep(): in the case when 1014 * the source page gets split and a tail freed before copy completes. 1015 * Called under pmd_lock of checked pmd, so safe from splitting itself. 1016 */ 1017 static void get_user_huge_page(struct page *page) 1018 { 1019 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) { 1020 struct page *endpage = page + HPAGE_PMD_NR; 1021 1022 atomic_add(HPAGE_PMD_NR, &page->_count); 1023 while (++page < endpage) 1024 get_huge_page_tail(page); 1025 } else { 1026 get_page(page); 1027 } 1028 } 1029 1030 static void put_user_huge_page(struct page *page) 1031 { 1032 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) { 1033 struct page *endpage = page + HPAGE_PMD_NR; 1034 1035 while (page < endpage) 1036 put_page(page++); 1037 } else { 1038 put_page(page); 1039 } 1040 } 1041 1042 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm, 1043 struct vm_area_struct *vma, 1044 unsigned long address, 1045 pmd_t *pmd, pmd_t orig_pmd, 1046 struct page *page, 1047 unsigned long haddr) 1048 { 1049 struct mem_cgroup *memcg; 1050 spinlock_t *ptl; 1051 pgtable_t pgtable; 1052 pmd_t _pmd; 1053 int ret = 0, i; 1054 struct page **pages; 1055 unsigned long mmun_start; /* For mmu_notifiers */ 1056 unsigned long mmun_end; /* For mmu_notifiers */ 1057 1058 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR, 1059 GFP_KERNEL); 1060 if (unlikely(!pages)) { 1061 ret |= VM_FAULT_OOM; 1062 goto out; 1063 } 1064 1065 for (i = 0; i < HPAGE_PMD_NR; i++) { 1066 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE | 1067 __GFP_OTHER_NODE, 1068 vma, address, page_to_nid(page)); 1069 if (unlikely(!pages[i] || 1070 mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL, 1071 &memcg))) { 1072 if (pages[i]) 1073 put_page(pages[i]); 1074 while (--i >= 0) { 1075 memcg = (void *)page_private(pages[i]); 1076 set_page_private(pages[i], 0); 1077 mem_cgroup_cancel_charge(pages[i], memcg); 1078 put_page(pages[i]); 1079 } 1080 kfree(pages); 1081 ret |= VM_FAULT_OOM; 1082 goto out; 1083 } 1084 set_page_private(pages[i], (unsigned long)memcg); 1085 } 1086 1087 for (i = 0; i < HPAGE_PMD_NR; i++) { 1088 copy_user_highpage(pages[i], page + i, 1089 haddr + PAGE_SIZE * i, vma); 1090 __SetPageUptodate(pages[i]); 1091 cond_resched(); 1092 } 1093 1094 mmun_start = haddr; 1095 mmun_end = haddr + HPAGE_PMD_SIZE; 1096 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 1097 1098 ptl = pmd_lock(mm, pmd); 1099 if (unlikely(!pmd_same(*pmd, orig_pmd))) 1100 goto out_free_pages; 1101 VM_BUG_ON_PAGE(!PageHead(page), page); 1102 1103 pmdp_huge_clear_flush_notify(vma, haddr, pmd); 1104 /* leave pmd empty until pte is filled */ 1105 1106 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 1107 pmd_populate(mm, &_pmd, pgtable); 1108 1109 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 1110 pte_t *pte, entry; 1111 entry = mk_pte(pages[i], vma->vm_page_prot); 1112 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 1113 memcg = (void *)page_private(pages[i]); 1114 set_page_private(pages[i], 0); 1115 page_add_new_anon_rmap(pages[i], vma, haddr); 1116 mem_cgroup_commit_charge(pages[i], memcg, false); 1117 lru_cache_add_active_or_unevictable(pages[i], vma); 1118 pte = pte_offset_map(&_pmd, haddr); 1119 VM_BUG_ON(!pte_none(*pte)); 1120 set_pte_at(mm, haddr, pte, entry); 1121 pte_unmap(pte); 1122 } 1123 kfree(pages); 1124 1125 smp_wmb(); /* make pte visible before pmd */ 1126 pmd_populate(mm, pmd, pgtable); 1127 page_remove_rmap(page); 1128 spin_unlock(ptl); 1129 1130 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 1131 1132 ret |= VM_FAULT_WRITE; 1133 put_page(page); 1134 1135 out: 1136 return ret; 1137 1138 out_free_pages: 1139 spin_unlock(ptl); 1140 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 1141 for (i = 0; i < HPAGE_PMD_NR; i++) { 1142 memcg = (void *)page_private(pages[i]); 1143 set_page_private(pages[i], 0); 1144 mem_cgroup_cancel_charge(pages[i], memcg); 1145 put_page(pages[i]); 1146 } 1147 kfree(pages); 1148 goto out; 1149 } 1150 1151 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma, 1152 unsigned long address, pmd_t *pmd, pmd_t orig_pmd) 1153 { 1154 spinlock_t *ptl; 1155 int ret = 0; 1156 struct page *page = NULL, *new_page; 1157 struct mem_cgroup *memcg; 1158 unsigned long haddr; 1159 unsigned long mmun_start; /* For mmu_notifiers */ 1160 unsigned long mmun_end; /* For mmu_notifiers */ 1161 gfp_t huge_gfp; /* for allocation and charge */ 1162 1163 ptl = pmd_lockptr(mm, pmd); 1164 VM_BUG_ON_VMA(!vma->anon_vma, vma); 1165 haddr = address & HPAGE_PMD_MASK; 1166 if (is_huge_zero_pmd(orig_pmd)) 1167 goto alloc; 1168 spin_lock(ptl); 1169 if (unlikely(!pmd_same(*pmd, orig_pmd))) 1170 goto out_unlock; 1171 1172 page = pmd_page(orig_pmd); 1173 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page); 1174 if (page_mapcount(page) == 1) { 1175 pmd_t entry; 1176 entry = pmd_mkyoung(orig_pmd); 1177 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 1178 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1)) 1179 update_mmu_cache_pmd(vma, address, pmd); 1180 ret |= VM_FAULT_WRITE; 1181 goto out_unlock; 1182 } 1183 get_user_huge_page(page); 1184 spin_unlock(ptl); 1185 alloc: 1186 if (transparent_hugepage_enabled(vma) && 1187 !transparent_hugepage_debug_cow()) { 1188 huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0); 1189 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER); 1190 } else 1191 new_page = NULL; 1192 1193 if (unlikely(!new_page)) { 1194 if (!page) { 1195 split_huge_page_pmd(vma, address, pmd); 1196 ret |= VM_FAULT_FALLBACK; 1197 } else { 1198 ret = do_huge_pmd_wp_page_fallback(mm, vma, address, 1199 pmd, orig_pmd, page, haddr); 1200 if (ret & VM_FAULT_OOM) { 1201 split_huge_page(page); 1202 ret |= VM_FAULT_FALLBACK; 1203 } 1204 put_user_huge_page(page); 1205 } 1206 count_vm_event(THP_FAULT_FALLBACK); 1207 goto out; 1208 } 1209 1210 if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg))) { 1211 put_page(new_page); 1212 if (page) { 1213 split_huge_page(page); 1214 put_user_huge_page(page); 1215 } else 1216 split_huge_page_pmd(vma, address, pmd); 1217 ret |= VM_FAULT_FALLBACK; 1218 count_vm_event(THP_FAULT_FALLBACK); 1219 goto out; 1220 } 1221 1222 count_vm_event(THP_FAULT_ALLOC); 1223 1224 if (!page) 1225 clear_huge_page(new_page, haddr, HPAGE_PMD_NR); 1226 else 1227 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR); 1228 __SetPageUptodate(new_page); 1229 1230 mmun_start = haddr; 1231 mmun_end = haddr + HPAGE_PMD_SIZE; 1232 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 1233 1234 spin_lock(ptl); 1235 if (page) 1236 put_user_huge_page(page); 1237 if (unlikely(!pmd_same(*pmd, orig_pmd))) { 1238 spin_unlock(ptl); 1239 mem_cgroup_cancel_charge(new_page, memcg); 1240 put_page(new_page); 1241 goto out_mn; 1242 } else { 1243 pmd_t entry; 1244 entry = mk_huge_pmd(new_page, vma->vm_page_prot); 1245 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 1246 pmdp_huge_clear_flush_notify(vma, haddr, pmd); 1247 page_add_new_anon_rmap(new_page, vma, haddr); 1248 mem_cgroup_commit_charge(new_page, memcg, false); 1249 lru_cache_add_active_or_unevictable(new_page, vma); 1250 set_pmd_at(mm, haddr, pmd, entry); 1251 update_mmu_cache_pmd(vma, address, pmd); 1252 if (!page) { 1253 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR); 1254 put_huge_zero_page(); 1255 } else { 1256 VM_BUG_ON_PAGE(!PageHead(page), page); 1257 page_remove_rmap(page); 1258 put_page(page); 1259 } 1260 ret |= VM_FAULT_WRITE; 1261 } 1262 spin_unlock(ptl); 1263 out_mn: 1264 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 1265 out: 1266 return ret; 1267 out_unlock: 1268 spin_unlock(ptl); 1269 return ret; 1270 } 1271 1272 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, 1273 unsigned long addr, 1274 pmd_t *pmd, 1275 unsigned int flags) 1276 { 1277 struct mm_struct *mm = vma->vm_mm; 1278 struct page *page = NULL; 1279 1280 assert_spin_locked(pmd_lockptr(mm, pmd)); 1281 1282 if (flags & FOLL_WRITE && !pmd_write(*pmd)) 1283 goto out; 1284 1285 /* Avoid dumping huge zero page */ 1286 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd)) 1287 return ERR_PTR(-EFAULT); 1288 1289 /* Full NUMA hinting faults to serialise migration in fault paths */ 1290 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd)) 1291 goto out; 1292 1293 page = pmd_page(*pmd); 1294 VM_BUG_ON_PAGE(!PageHead(page), page); 1295 if (flags & FOLL_TOUCH) { 1296 pmd_t _pmd; 1297 /* 1298 * We should set the dirty bit only for FOLL_WRITE but 1299 * for now the dirty bit in the pmd is meaningless. 1300 * And if the dirty bit will become meaningful and 1301 * we'll only set it with FOLL_WRITE, an atomic 1302 * set_bit will be required on the pmd to set the 1303 * young bit, instead of the current set_pmd_at. 1304 */ 1305 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd)); 1306 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK, 1307 pmd, _pmd, 1)) 1308 update_mmu_cache_pmd(vma, addr, pmd); 1309 } 1310 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) { 1311 if (page->mapping && trylock_page(page)) { 1312 lru_add_drain(); 1313 if (page->mapping) 1314 mlock_vma_page(page); 1315 unlock_page(page); 1316 } 1317 } 1318 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; 1319 VM_BUG_ON_PAGE(!PageCompound(page), page); 1320 if (flags & FOLL_GET) 1321 get_page_foll(page); 1322 1323 out: 1324 return page; 1325 } 1326 1327 /* NUMA hinting page fault entry point for trans huge pmds */ 1328 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma, 1329 unsigned long addr, pmd_t pmd, pmd_t *pmdp) 1330 { 1331 spinlock_t *ptl; 1332 struct anon_vma *anon_vma = NULL; 1333 struct page *page; 1334 unsigned long haddr = addr & HPAGE_PMD_MASK; 1335 int page_nid = -1, this_nid = numa_node_id(); 1336 int target_nid, last_cpupid = -1; 1337 bool page_locked; 1338 bool migrated = false; 1339 bool was_writable; 1340 int flags = 0; 1341 1342 /* A PROT_NONE fault should not end up here */ 1343 BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))); 1344 1345 ptl = pmd_lock(mm, pmdp); 1346 if (unlikely(!pmd_same(pmd, *pmdp))) 1347 goto out_unlock; 1348 1349 /* 1350 * If there are potential migrations, wait for completion and retry 1351 * without disrupting NUMA hinting information. Do not relock and 1352 * check_same as the page may no longer be mapped. 1353 */ 1354 if (unlikely(pmd_trans_migrating(*pmdp))) { 1355 page = pmd_page(*pmdp); 1356 spin_unlock(ptl); 1357 wait_on_page_locked(page); 1358 goto out; 1359 } 1360 1361 page = pmd_page(pmd); 1362 BUG_ON(is_huge_zero_page(page)); 1363 page_nid = page_to_nid(page); 1364 last_cpupid = page_cpupid_last(page); 1365 count_vm_numa_event(NUMA_HINT_FAULTS); 1366 if (page_nid == this_nid) { 1367 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); 1368 flags |= TNF_FAULT_LOCAL; 1369 } 1370 1371 /* See similar comment in do_numa_page for explanation */ 1372 if (!(vma->vm_flags & VM_WRITE)) 1373 flags |= TNF_NO_GROUP; 1374 1375 /* 1376 * Acquire the page lock to serialise THP migrations but avoid dropping 1377 * page_table_lock if at all possible 1378 */ 1379 page_locked = trylock_page(page); 1380 target_nid = mpol_misplaced(page, vma, haddr); 1381 if (target_nid == -1) { 1382 /* If the page was locked, there are no parallel migrations */ 1383 if (page_locked) 1384 goto clear_pmdnuma; 1385 } 1386 1387 /* Migration could have started since the pmd_trans_migrating check */ 1388 if (!page_locked) { 1389 spin_unlock(ptl); 1390 wait_on_page_locked(page); 1391 page_nid = -1; 1392 goto out; 1393 } 1394 1395 /* 1396 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma 1397 * to serialises splits 1398 */ 1399 get_page(page); 1400 spin_unlock(ptl); 1401 anon_vma = page_lock_anon_vma_read(page); 1402 1403 /* Confirm the PMD did not change while page_table_lock was released */ 1404 spin_lock(ptl); 1405 if (unlikely(!pmd_same(pmd, *pmdp))) { 1406 unlock_page(page); 1407 put_page(page); 1408 page_nid = -1; 1409 goto out_unlock; 1410 } 1411 1412 /* Bail if we fail to protect against THP splits for any reason */ 1413 if (unlikely(!anon_vma)) { 1414 put_page(page); 1415 page_nid = -1; 1416 goto clear_pmdnuma; 1417 } 1418 1419 /* 1420 * Migrate the THP to the requested node, returns with page unlocked 1421 * and access rights restored. 1422 */ 1423 spin_unlock(ptl); 1424 migrated = migrate_misplaced_transhuge_page(mm, vma, 1425 pmdp, pmd, addr, page, target_nid); 1426 if (migrated) { 1427 flags |= TNF_MIGRATED; 1428 page_nid = target_nid; 1429 } else 1430 flags |= TNF_MIGRATE_FAIL; 1431 1432 goto out; 1433 clear_pmdnuma: 1434 BUG_ON(!PageLocked(page)); 1435 was_writable = pmd_write(pmd); 1436 pmd = pmd_modify(pmd, vma->vm_page_prot); 1437 pmd = pmd_mkyoung(pmd); 1438 if (was_writable) 1439 pmd = pmd_mkwrite(pmd); 1440 set_pmd_at(mm, haddr, pmdp, pmd); 1441 update_mmu_cache_pmd(vma, addr, pmdp); 1442 unlock_page(page); 1443 out_unlock: 1444 spin_unlock(ptl); 1445 1446 out: 1447 if (anon_vma) 1448 page_unlock_anon_vma_read(anon_vma); 1449 1450 if (page_nid != -1) 1451 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags); 1452 1453 return 0; 1454 } 1455 1456 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1457 pmd_t *pmd, unsigned long addr) 1458 { 1459 pmd_t orig_pmd; 1460 spinlock_t *ptl; 1461 1462 if (__pmd_trans_huge_lock(pmd, vma, &ptl) != 1) 1463 return 0; 1464 /* 1465 * For architectures like ppc64 we look at deposited pgtable 1466 * when calling pmdp_huge_get_and_clear. So do the 1467 * pgtable_trans_huge_withdraw after finishing pmdp related 1468 * operations. 1469 */ 1470 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd, 1471 tlb->fullmm); 1472 tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 1473 if (vma_is_dax(vma)) { 1474 spin_unlock(ptl); 1475 if (is_huge_zero_pmd(orig_pmd)) 1476 put_huge_zero_page(); 1477 } else if (is_huge_zero_pmd(orig_pmd)) { 1478 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd)); 1479 atomic_long_dec(&tlb->mm->nr_ptes); 1480 spin_unlock(ptl); 1481 put_huge_zero_page(); 1482 } else { 1483 struct page *page = pmd_page(orig_pmd); 1484 page_remove_rmap(page); 1485 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page); 1486 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); 1487 VM_BUG_ON_PAGE(!PageHead(page), page); 1488 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd)); 1489 atomic_long_dec(&tlb->mm->nr_ptes); 1490 spin_unlock(ptl); 1491 tlb_remove_page(tlb, page); 1492 } 1493 return 1; 1494 } 1495 1496 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma, 1497 unsigned long old_addr, 1498 unsigned long new_addr, unsigned long old_end, 1499 pmd_t *old_pmd, pmd_t *new_pmd) 1500 { 1501 spinlock_t *old_ptl, *new_ptl; 1502 int ret = 0; 1503 pmd_t pmd; 1504 1505 struct mm_struct *mm = vma->vm_mm; 1506 1507 if ((old_addr & ~HPAGE_PMD_MASK) || 1508 (new_addr & ~HPAGE_PMD_MASK) || 1509 old_end - old_addr < HPAGE_PMD_SIZE || 1510 (new_vma->vm_flags & VM_NOHUGEPAGE)) 1511 goto out; 1512 1513 /* 1514 * The destination pmd shouldn't be established, free_pgtables() 1515 * should have release it. 1516 */ 1517 if (WARN_ON(!pmd_none(*new_pmd))) { 1518 VM_BUG_ON(pmd_trans_huge(*new_pmd)); 1519 goto out; 1520 } 1521 1522 /* 1523 * We don't have to worry about the ordering of src and dst 1524 * ptlocks because exclusive mmap_sem prevents deadlock. 1525 */ 1526 ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl); 1527 if (ret == 1) { 1528 new_ptl = pmd_lockptr(mm, new_pmd); 1529 if (new_ptl != old_ptl) 1530 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); 1531 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd); 1532 VM_BUG_ON(!pmd_none(*new_pmd)); 1533 1534 if (pmd_move_must_withdraw(new_ptl, old_ptl)) { 1535 pgtable_t pgtable; 1536 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd); 1537 pgtable_trans_huge_deposit(mm, new_pmd, pgtable); 1538 } 1539 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd)); 1540 if (new_ptl != old_ptl) 1541 spin_unlock(new_ptl); 1542 spin_unlock(old_ptl); 1543 } 1544 out: 1545 return ret; 1546 } 1547 1548 /* 1549 * Returns 1550 * - 0 if PMD could not be locked 1551 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary 1552 * - HPAGE_PMD_NR is protections changed and TLB flush necessary 1553 */ 1554 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 1555 unsigned long addr, pgprot_t newprot, int prot_numa) 1556 { 1557 struct mm_struct *mm = vma->vm_mm; 1558 spinlock_t *ptl; 1559 int ret = 0; 1560 1561 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) { 1562 pmd_t entry; 1563 bool preserve_write = prot_numa && pmd_write(*pmd); 1564 ret = 1; 1565 1566 /* 1567 * Avoid trapping faults against the zero page. The read-only 1568 * data is likely to be read-cached on the local CPU and 1569 * local/remote hits to the zero page are not interesting. 1570 */ 1571 if (prot_numa && is_huge_zero_pmd(*pmd)) { 1572 spin_unlock(ptl); 1573 return ret; 1574 } 1575 1576 if (!prot_numa || !pmd_protnone(*pmd)) { 1577 entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd); 1578 entry = pmd_modify(entry, newprot); 1579 if (preserve_write) 1580 entry = pmd_mkwrite(entry); 1581 ret = HPAGE_PMD_NR; 1582 set_pmd_at(mm, addr, pmd, entry); 1583 BUG_ON(!preserve_write && pmd_write(entry)); 1584 } 1585 spin_unlock(ptl); 1586 } 1587 1588 return ret; 1589 } 1590 1591 /* 1592 * Returns 1 if a given pmd maps a stable (not under splitting) thp. 1593 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise. 1594 * 1595 * Note that if it returns 1, this routine returns without unlocking page 1596 * table locks. So callers must unlock them. 1597 */ 1598 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma, 1599 spinlock_t **ptl) 1600 { 1601 *ptl = pmd_lock(vma->vm_mm, pmd); 1602 if (likely(pmd_trans_huge(*pmd))) { 1603 if (unlikely(pmd_trans_splitting(*pmd))) { 1604 spin_unlock(*ptl); 1605 wait_split_huge_page(vma->anon_vma, pmd); 1606 return -1; 1607 } else { 1608 /* Thp mapped by 'pmd' is stable, so we can 1609 * handle it as it is. */ 1610 return 1; 1611 } 1612 } 1613 spin_unlock(*ptl); 1614 return 0; 1615 } 1616 1617 /* 1618 * This function returns whether a given @page is mapped onto the @address 1619 * in the virtual space of @mm. 1620 * 1621 * When it's true, this function returns *pmd with holding the page table lock 1622 * and passing it back to the caller via @ptl. 1623 * If it's false, returns NULL without holding the page table lock. 1624 */ 1625 pmd_t *page_check_address_pmd(struct page *page, 1626 struct mm_struct *mm, 1627 unsigned long address, 1628 enum page_check_address_pmd_flag flag, 1629 spinlock_t **ptl) 1630 { 1631 pgd_t *pgd; 1632 pud_t *pud; 1633 pmd_t *pmd; 1634 1635 if (address & ~HPAGE_PMD_MASK) 1636 return NULL; 1637 1638 pgd = pgd_offset(mm, address); 1639 if (!pgd_present(*pgd)) 1640 return NULL; 1641 pud = pud_offset(pgd, address); 1642 if (!pud_present(*pud)) 1643 return NULL; 1644 pmd = pmd_offset(pud, address); 1645 1646 *ptl = pmd_lock(mm, pmd); 1647 if (!pmd_present(*pmd)) 1648 goto unlock; 1649 if (pmd_page(*pmd) != page) 1650 goto unlock; 1651 /* 1652 * split_vma() may create temporary aliased mappings. There is 1653 * no risk as long as all huge pmd are found and have their 1654 * splitting bit set before __split_huge_page_refcount 1655 * runs. Finding the same huge pmd more than once during the 1656 * same rmap walk is not a problem. 1657 */ 1658 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG && 1659 pmd_trans_splitting(*pmd)) 1660 goto unlock; 1661 if (pmd_trans_huge(*pmd)) { 1662 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG && 1663 !pmd_trans_splitting(*pmd)); 1664 return pmd; 1665 } 1666 unlock: 1667 spin_unlock(*ptl); 1668 return NULL; 1669 } 1670 1671 static int __split_huge_page_splitting(struct page *page, 1672 struct vm_area_struct *vma, 1673 unsigned long address) 1674 { 1675 struct mm_struct *mm = vma->vm_mm; 1676 spinlock_t *ptl; 1677 pmd_t *pmd; 1678 int ret = 0; 1679 /* For mmu_notifiers */ 1680 const unsigned long mmun_start = address; 1681 const unsigned long mmun_end = address + HPAGE_PMD_SIZE; 1682 1683 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 1684 pmd = page_check_address_pmd(page, mm, address, 1685 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl); 1686 if (pmd) { 1687 /* 1688 * We can't temporarily set the pmd to null in order 1689 * to split it, the pmd must remain marked huge at all 1690 * times or the VM won't take the pmd_trans_huge paths 1691 * and it won't wait on the anon_vma->root->rwsem to 1692 * serialize against split_huge_page*. 1693 */ 1694 pmdp_splitting_flush(vma, address, pmd); 1695 1696 ret = 1; 1697 spin_unlock(ptl); 1698 } 1699 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 1700 1701 return ret; 1702 } 1703 1704 static void __split_huge_page_refcount(struct page *page, 1705 struct list_head *list) 1706 { 1707 int i; 1708 struct zone *zone = page_zone(page); 1709 struct lruvec *lruvec; 1710 int tail_count = 0; 1711 1712 /* prevent PageLRU to go away from under us, and freeze lru stats */ 1713 spin_lock_irq(&zone->lru_lock); 1714 lruvec = mem_cgroup_page_lruvec(page, zone); 1715 1716 compound_lock(page); 1717 /* complete memcg works before add pages to LRU */ 1718 mem_cgroup_split_huge_fixup(page); 1719 1720 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) { 1721 struct page *page_tail = page + i; 1722 1723 /* tail_page->_mapcount cannot change */ 1724 BUG_ON(page_mapcount(page_tail) < 0); 1725 tail_count += page_mapcount(page_tail); 1726 /* check for overflow */ 1727 BUG_ON(tail_count < 0); 1728 BUG_ON(atomic_read(&page_tail->_count) != 0); 1729 /* 1730 * tail_page->_count is zero and not changing from 1731 * under us. But get_page_unless_zero() may be running 1732 * from under us on the tail_page. If we used 1733 * atomic_set() below instead of atomic_add(), we 1734 * would then run atomic_set() concurrently with 1735 * get_page_unless_zero(), and atomic_set() is 1736 * implemented in C not using locked ops. spin_unlock 1737 * on x86 sometime uses locked ops because of PPro 1738 * errata 66, 92, so unless somebody can guarantee 1739 * atomic_set() here would be safe on all archs (and 1740 * not only on x86), it's safer to use atomic_add(). 1741 */ 1742 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1, 1743 &page_tail->_count); 1744 1745 /* after clearing PageTail the gup refcount can be released */ 1746 smp_mb__after_atomic(); 1747 1748 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 1749 page_tail->flags |= (page->flags & 1750 ((1L << PG_referenced) | 1751 (1L << PG_swapbacked) | 1752 (1L << PG_mlocked) | 1753 (1L << PG_uptodate) | 1754 (1L << PG_active) | 1755 (1L << PG_unevictable))); 1756 page_tail->flags |= (1L << PG_dirty); 1757 1758 clear_compound_head(page_tail); 1759 1760 if (page_is_young(page)) 1761 set_page_young(page_tail); 1762 if (page_is_idle(page)) 1763 set_page_idle(page_tail); 1764 1765 /* 1766 * __split_huge_page_splitting() already set the 1767 * splitting bit in all pmd that could map this 1768 * hugepage, that will ensure no CPU can alter the 1769 * mapcount on the head page. The mapcount is only 1770 * accounted in the head page and it has to be 1771 * transferred to all tail pages in the below code. So 1772 * for this code to be safe, the split the mapcount 1773 * can't change. But that doesn't mean userland can't 1774 * keep changing and reading the page contents while 1775 * we transfer the mapcount, so the pmd splitting 1776 * status is achieved setting a reserved bit in the 1777 * pmd, not by clearing the present bit. 1778 */ 1779 page_tail->_mapcount = page->_mapcount; 1780 1781 BUG_ON(page_tail->mapping); 1782 page_tail->mapping = page->mapping; 1783 1784 page_tail->index = page->index + i; 1785 page_cpupid_xchg_last(page_tail, page_cpupid_last(page)); 1786 1787 BUG_ON(!PageAnon(page_tail)); 1788 BUG_ON(!PageUptodate(page_tail)); 1789 BUG_ON(!PageDirty(page_tail)); 1790 BUG_ON(!PageSwapBacked(page_tail)); 1791 1792 lru_add_page_tail(page, page_tail, lruvec, list); 1793 } 1794 atomic_sub(tail_count, &page->_count); 1795 BUG_ON(atomic_read(&page->_count) <= 0); 1796 1797 __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1); 1798 1799 ClearPageCompound(page); 1800 compound_unlock(page); 1801 spin_unlock_irq(&zone->lru_lock); 1802 1803 for (i = 1; i < HPAGE_PMD_NR; i++) { 1804 struct page *page_tail = page + i; 1805 BUG_ON(page_count(page_tail) <= 0); 1806 /* 1807 * Tail pages may be freed if there wasn't any mapping 1808 * like if add_to_swap() is running on a lru page that 1809 * had its mapping zapped. And freeing these pages 1810 * requires taking the lru_lock so we do the put_page 1811 * of the tail pages after the split is complete. 1812 */ 1813 put_page(page_tail); 1814 } 1815 1816 /* 1817 * Only the head page (now become a regular page) is required 1818 * to be pinned by the caller. 1819 */ 1820 BUG_ON(page_count(page) <= 0); 1821 } 1822 1823 static int __split_huge_page_map(struct page *page, 1824 struct vm_area_struct *vma, 1825 unsigned long address) 1826 { 1827 struct mm_struct *mm = vma->vm_mm; 1828 spinlock_t *ptl; 1829 pmd_t *pmd, _pmd; 1830 int ret = 0, i; 1831 pgtable_t pgtable; 1832 unsigned long haddr; 1833 1834 pmd = page_check_address_pmd(page, mm, address, 1835 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl); 1836 if (pmd) { 1837 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 1838 pmd_populate(mm, &_pmd, pgtable); 1839 if (pmd_write(*pmd)) 1840 BUG_ON(page_mapcount(page) != 1); 1841 1842 haddr = address; 1843 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 1844 pte_t *pte, entry; 1845 BUG_ON(PageCompound(page+i)); 1846 /* 1847 * Note that NUMA hinting access restrictions are not 1848 * transferred to avoid any possibility of altering 1849 * permissions across VMAs. 1850 */ 1851 entry = mk_pte(page + i, vma->vm_page_prot); 1852 entry = maybe_mkwrite(pte_mkdirty(entry), vma); 1853 if (!pmd_write(*pmd)) 1854 entry = pte_wrprotect(entry); 1855 if (!pmd_young(*pmd)) 1856 entry = pte_mkold(entry); 1857 pte = pte_offset_map(&_pmd, haddr); 1858 BUG_ON(!pte_none(*pte)); 1859 set_pte_at(mm, haddr, pte, entry); 1860 pte_unmap(pte); 1861 } 1862 1863 smp_wmb(); /* make pte visible before pmd */ 1864 /* 1865 * Up to this point the pmd is present and huge and 1866 * userland has the whole access to the hugepage 1867 * during the split (which happens in place). If we 1868 * overwrite the pmd with the not-huge version 1869 * pointing to the pte here (which of course we could 1870 * if all CPUs were bug free), userland could trigger 1871 * a small page size TLB miss on the small sized TLB 1872 * while the hugepage TLB entry is still established 1873 * in the huge TLB. Some CPU doesn't like that. See 1874 * http://support.amd.com/us/Processor_TechDocs/41322.pdf, 1875 * Erratum 383 on page 93. Intel should be safe but is 1876 * also warns that it's only safe if the permission 1877 * and cache attributes of the two entries loaded in 1878 * the two TLB is identical (which should be the case 1879 * here). But it is generally safer to never allow 1880 * small and huge TLB entries for the same virtual 1881 * address to be loaded simultaneously. So instead of 1882 * doing "pmd_populate(); flush_pmd_tlb_range();" we first 1883 * mark the current pmd notpresent (atomically because 1884 * here the pmd_trans_huge and pmd_trans_splitting 1885 * must remain set at all times on the pmd until the 1886 * split is complete for this pmd), then we flush the 1887 * SMP TLB and finally we write the non-huge version 1888 * of the pmd entry with pmd_populate. 1889 */ 1890 pmdp_invalidate(vma, address, pmd); 1891 pmd_populate(mm, pmd, pgtable); 1892 ret = 1; 1893 spin_unlock(ptl); 1894 } 1895 1896 return ret; 1897 } 1898 1899 /* must be called with anon_vma->root->rwsem held */ 1900 static void __split_huge_page(struct page *page, 1901 struct anon_vma *anon_vma, 1902 struct list_head *list) 1903 { 1904 int mapcount, mapcount2; 1905 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); 1906 struct anon_vma_chain *avc; 1907 1908 BUG_ON(!PageHead(page)); 1909 BUG_ON(PageTail(page)); 1910 1911 mapcount = 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 mapcount += __split_huge_page_splitting(page, vma, addr); 1917 } 1918 /* 1919 * It is critical that new vmas are added to the tail of the 1920 * anon_vma list. This guarantes that if copy_huge_pmd() runs 1921 * and establishes a child pmd before 1922 * __split_huge_page_splitting() freezes the parent pmd (so if 1923 * we fail to prevent copy_huge_pmd() from running until the 1924 * whole __split_huge_page() is complete), we will still see 1925 * the newly established pmd of the child later during the 1926 * walk, to be able to set it as pmd_trans_splitting too. 1927 */ 1928 if (mapcount != page_mapcount(page)) { 1929 pr_err("mapcount %d page_mapcount %d\n", 1930 mapcount, page_mapcount(page)); 1931 BUG(); 1932 } 1933 1934 __split_huge_page_refcount(page, list); 1935 1936 mapcount2 = 0; 1937 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) { 1938 struct vm_area_struct *vma = avc->vma; 1939 unsigned long addr = vma_address(page, vma); 1940 BUG_ON(is_vma_temporary_stack(vma)); 1941 mapcount2 += __split_huge_page_map(page, vma, addr); 1942 } 1943 if (mapcount != mapcount2) { 1944 pr_err("mapcount %d mapcount2 %d page_mapcount %d\n", 1945 mapcount, mapcount2, page_mapcount(page)); 1946 BUG(); 1947 } 1948 } 1949 1950 /* 1951 * Split a hugepage into normal pages. This doesn't change the position of head 1952 * page. If @list is null, tail pages will be added to LRU list, otherwise, to 1953 * @list. Both head page and tail pages will inherit mapping, flags, and so on 1954 * from the hugepage. 1955 * Return 0 if the hugepage is split successfully otherwise return 1. 1956 */ 1957 int split_huge_page_to_list(struct page *page, struct list_head *list) 1958 { 1959 struct anon_vma *anon_vma; 1960 int ret = 1; 1961 1962 BUG_ON(is_huge_zero_page(page)); 1963 BUG_ON(!PageAnon(page)); 1964 1965 /* 1966 * The caller does not necessarily hold an mmap_sem that would prevent 1967 * the anon_vma disappearing so we first we take a reference to it 1968 * and then lock the anon_vma for write. This is similar to 1969 * page_lock_anon_vma_read except the write lock is taken to serialise 1970 * against parallel split or collapse operations. 1971 */ 1972 anon_vma = page_get_anon_vma(page); 1973 if (!anon_vma) 1974 goto out; 1975 anon_vma_lock_write(anon_vma); 1976 1977 ret = 0; 1978 if (!PageCompound(page)) 1979 goto out_unlock; 1980 1981 BUG_ON(!PageSwapBacked(page)); 1982 __split_huge_page(page, anon_vma, list); 1983 count_vm_event(THP_SPLIT); 1984 1985 BUG_ON(PageCompound(page)); 1986 out_unlock: 1987 anon_vma_unlock_write(anon_vma); 1988 put_anon_vma(anon_vma); 1989 out: 1990 return ret; 1991 } 1992 1993 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE) 1994 1995 int hugepage_madvise(struct vm_area_struct *vma, 1996 unsigned long *vm_flags, int advice) 1997 { 1998 switch (advice) { 1999 case MADV_HUGEPAGE: 2000 #ifdef CONFIG_S390 2001 /* 2002 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390 2003 * can't handle this properly after s390_enable_sie, so we simply 2004 * ignore the madvise to prevent qemu from causing a SIGSEGV. 2005 */ 2006 if (mm_has_pgste(vma->vm_mm)) 2007 return 0; 2008 #endif 2009 /* 2010 * Be somewhat over-protective like KSM for now! 2011 */ 2012 if (*vm_flags & VM_NO_THP) 2013 return -EINVAL; 2014 *vm_flags &= ~VM_NOHUGEPAGE; 2015 *vm_flags |= VM_HUGEPAGE; 2016 /* 2017 * If the vma become good for khugepaged to scan, 2018 * register it here without waiting a page fault that 2019 * may not happen any time soon. 2020 */ 2021 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags))) 2022 return -ENOMEM; 2023 break; 2024 case MADV_NOHUGEPAGE: 2025 /* 2026 * Be somewhat over-protective like KSM for now! 2027 */ 2028 if (*vm_flags & VM_NO_THP) 2029 return -EINVAL; 2030 *vm_flags &= ~VM_HUGEPAGE; 2031 *vm_flags |= VM_NOHUGEPAGE; 2032 /* 2033 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning 2034 * this vma even if we leave the mm registered in khugepaged if 2035 * it got registered before VM_NOHUGEPAGE was set. 2036 */ 2037 break; 2038 } 2039 2040 return 0; 2041 } 2042 2043 static int __init khugepaged_slab_init(void) 2044 { 2045 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot", 2046 sizeof(struct mm_slot), 2047 __alignof__(struct mm_slot), 0, NULL); 2048 if (!mm_slot_cache) 2049 return -ENOMEM; 2050 2051 return 0; 2052 } 2053 2054 static void __init khugepaged_slab_exit(void) 2055 { 2056 kmem_cache_destroy(mm_slot_cache); 2057 } 2058 2059 static inline struct mm_slot *alloc_mm_slot(void) 2060 { 2061 if (!mm_slot_cache) /* initialization failed */ 2062 return NULL; 2063 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL); 2064 } 2065 2066 static inline void free_mm_slot(struct mm_slot *mm_slot) 2067 { 2068 kmem_cache_free(mm_slot_cache, mm_slot); 2069 } 2070 2071 static struct mm_slot *get_mm_slot(struct mm_struct *mm) 2072 { 2073 struct mm_slot *mm_slot; 2074 2075 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm) 2076 if (mm == mm_slot->mm) 2077 return mm_slot; 2078 2079 return NULL; 2080 } 2081 2082 static void insert_to_mm_slots_hash(struct mm_struct *mm, 2083 struct mm_slot *mm_slot) 2084 { 2085 mm_slot->mm = mm; 2086 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm); 2087 } 2088 2089 static inline int khugepaged_test_exit(struct mm_struct *mm) 2090 { 2091 return atomic_read(&mm->mm_users) == 0; 2092 } 2093 2094 int __khugepaged_enter(struct mm_struct *mm) 2095 { 2096 struct mm_slot *mm_slot; 2097 int wakeup; 2098 2099 mm_slot = alloc_mm_slot(); 2100 if (!mm_slot) 2101 return -ENOMEM; 2102 2103 /* __khugepaged_exit() must not run from under us */ 2104 VM_BUG_ON_MM(khugepaged_test_exit(mm), mm); 2105 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) { 2106 free_mm_slot(mm_slot); 2107 return 0; 2108 } 2109 2110 spin_lock(&khugepaged_mm_lock); 2111 insert_to_mm_slots_hash(mm, mm_slot); 2112 /* 2113 * Insert just behind the scanning cursor, to let the area settle 2114 * down a little. 2115 */ 2116 wakeup = list_empty(&khugepaged_scan.mm_head); 2117 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head); 2118 spin_unlock(&khugepaged_mm_lock); 2119 2120 atomic_inc(&mm->mm_count); 2121 if (wakeup) 2122 wake_up_interruptible(&khugepaged_wait); 2123 2124 return 0; 2125 } 2126 2127 int khugepaged_enter_vma_merge(struct vm_area_struct *vma, 2128 unsigned long vm_flags) 2129 { 2130 unsigned long hstart, hend; 2131 if (!vma->anon_vma) 2132 /* 2133 * Not yet faulted in so we will register later in the 2134 * page fault if needed. 2135 */ 2136 return 0; 2137 if (vma->vm_ops) 2138 /* khugepaged not yet working on file or special mappings */ 2139 return 0; 2140 VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma); 2141 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; 2142 hend = vma->vm_end & HPAGE_PMD_MASK; 2143 if (hstart < hend) 2144 return khugepaged_enter(vma, vm_flags); 2145 return 0; 2146 } 2147 2148 void __khugepaged_exit(struct mm_struct *mm) 2149 { 2150 struct mm_slot *mm_slot; 2151 int free = 0; 2152 2153 spin_lock(&khugepaged_mm_lock); 2154 mm_slot = get_mm_slot(mm); 2155 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) { 2156 hash_del(&mm_slot->hash); 2157 list_del(&mm_slot->mm_node); 2158 free = 1; 2159 } 2160 spin_unlock(&khugepaged_mm_lock); 2161 2162 if (free) { 2163 clear_bit(MMF_VM_HUGEPAGE, &mm->flags); 2164 free_mm_slot(mm_slot); 2165 mmdrop(mm); 2166 } else if (mm_slot) { 2167 /* 2168 * This is required to serialize against 2169 * khugepaged_test_exit() (which is guaranteed to run 2170 * under mmap sem read mode). Stop here (after we 2171 * return all pagetables will be destroyed) until 2172 * khugepaged has finished working on the pagetables 2173 * under the mmap_sem. 2174 */ 2175 down_write(&mm->mmap_sem); 2176 up_write(&mm->mmap_sem); 2177 } 2178 } 2179 2180 static void release_pte_page(struct page *page) 2181 { 2182 /* 0 stands for page_is_file_cache(page) == false */ 2183 dec_zone_page_state(page, NR_ISOLATED_ANON + 0); 2184 unlock_page(page); 2185 putback_lru_page(page); 2186 } 2187 2188 static void release_pte_pages(pte_t *pte, pte_t *_pte) 2189 { 2190 while (--_pte >= pte) { 2191 pte_t pteval = *_pte; 2192 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval))) 2193 release_pte_page(pte_page(pteval)); 2194 } 2195 } 2196 2197 static int __collapse_huge_page_isolate(struct vm_area_struct *vma, 2198 unsigned long address, 2199 pte_t *pte) 2200 { 2201 struct page *page; 2202 pte_t *_pte; 2203 int none_or_zero = 0; 2204 bool referenced = false, writable = false; 2205 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; 2206 _pte++, address += PAGE_SIZE) { 2207 pte_t pteval = *_pte; 2208 if (pte_none(pteval) || (pte_present(pteval) && 2209 is_zero_pfn(pte_pfn(pteval)))) { 2210 if (!userfaultfd_armed(vma) && 2211 ++none_or_zero <= khugepaged_max_ptes_none) 2212 continue; 2213 else 2214 goto out; 2215 } 2216 if (!pte_present(pteval)) 2217 goto out; 2218 page = vm_normal_page(vma, address, pteval); 2219 if (unlikely(!page)) 2220 goto out; 2221 2222 VM_BUG_ON_PAGE(PageCompound(page), page); 2223 VM_BUG_ON_PAGE(!PageAnon(page), page); 2224 VM_BUG_ON_PAGE(!PageSwapBacked(page), page); 2225 2226 /* 2227 * We can do it before isolate_lru_page because the 2228 * page can't be freed from under us. NOTE: PG_lock 2229 * is needed to serialize against split_huge_page 2230 * when invoked from the VM. 2231 */ 2232 if (!trylock_page(page)) 2233 goto out; 2234 2235 /* 2236 * cannot use mapcount: can't collapse if there's a gup pin. 2237 * The page must only be referenced by the scanned process 2238 * and page swap cache. 2239 */ 2240 if (page_count(page) != 1 + !!PageSwapCache(page)) { 2241 unlock_page(page); 2242 goto out; 2243 } 2244 if (pte_write(pteval)) { 2245 writable = true; 2246 } else { 2247 if (PageSwapCache(page) && !reuse_swap_page(page)) { 2248 unlock_page(page); 2249 goto out; 2250 } 2251 /* 2252 * Page is not in the swap cache. It can be collapsed 2253 * into a THP. 2254 */ 2255 } 2256 2257 /* 2258 * Isolate the page to avoid collapsing an hugepage 2259 * currently in use by the VM. 2260 */ 2261 if (isolate_lru_page(page)) { 2262 unlock_page(page); 2263 goto out; 2264 } 2265 /* 0 stands for page_is_file_cache(page) == false */ 2266 inc_zone_page_state(page, NR_ISOLATED_ANON + 0); 2267 VM_BUG_ON_PAGE(!PageLocked(page), page); 2268 VM_BUG_ON_PAGE(PageLRU(page), page); 2269 2270 /* If there is no mapped pte young don't collapse the page */ 2271 if (pte_young(pteval) || 2272 page_is_young(page) || PageReferenced(page) || 2273 mmu_notifier_test_young(vma->vm_mm, address)) 2274 referenced = true; 2275 } 2276 if (likely(referenced && writable)) 2277 return 1; 2278 out: 2279 release_pte_pages(pte, _pte); 2280 return 0; 2281 } 2282 2283 static void __collapse_huge_page_copy(pte_t *pte, struct page *page, 2284 struct vm_area_struct *vma, 2285 unsigned long address, 2286 spinlock_t *ptl) 2287 { 2288 pte_t *_pte; 2289 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) { 2290 pte_t pteval = *_pte; 2291 struct page *src_page; 2292 2293 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { 2294 clear_user_highpage(page, address); 2295 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); 2296 if (is_zero_pfn(pte_pfn(pteval))) { 2297 /* 2298 * ptl mostly unnecessary. 2299 */ 2300 spin_lock(ptl); 2301 /* 2302 * paravirt calls inside pte_clear here are 2303 * superfluous. 2304 */ 2305 pte_clear(vma->vm_mm, address, _pte); 2306 spin_unlock(ptl); 2307 } 2308 } else { 2309 src_page = pte_page(pteval); 2310 copy_user_highpage(page, src_page, address, vma); 2311 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page); 2312 release_pte_page(src_page); 2313 /* 2314 * ptl mostly unnecessary, but preempt has to 2315 * be disabled to update the per-cpu stats 2316 * inside page_remove_rmap(). 2317 */ 2318 spin_lock(ptl); 2319 /* 2320 * paravirt calls inside pte_clear here are 2321 * superfluous. 2322 */ 2323 pte_clear(vma->vm_mm, address, _pte); 2324 page_remove_rmap(src_page); 2325 spin_unlock(ptl); 2326 free_page_and_swap_cache(src_page); 2327 } 2328 2329 address += PAGE_SIZE; 2330 page++; 2331 } 2332 } 2333 2334 static void khugepaged_alloc_sleep(void) 2335 { 2336 DEFINE_WAIT(wait); 2337 2338 add_wait_queue(&khugepaged_wait, &wait); 2339 freezable_schedule_timeout_interruptible( 2340 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs)); 2341 remove_wait_queue(&khugepaged_wait, &wait); 2342 } 2343 2344 static int khugepaged_node_load[MAX_NUMNODES]; 2345 2346 static bool khugepaged_scan_abort(int nid) 2347 { 2348 int i; 2349 2350 /* 2351 * If zone_reclaim_mode is disabled, then no extra effort is made to 2352 * allocate memory locally. 2353 */ 2354 if (!zone_reclaim_mode) 2355 return false; 2356 2357 /* If there is a count for this node already, it must be acceptable */ 2358 if (khugepaged_node_load[nid]) 2359 return false; 2360 2361 for (i = 0; i < MAX_NUMNODES; i++) { 2362 if (!khugepaged_node_load[i]) 2363 continue; 2364 if (node_distance(nid, i) > RECLAIM_DISTANCE) 2365 return true; 2366 } 2367 return false; 2368 } 2369 2370 #ifdef CONFIG_NUMA 2371 static int khugepaged_find_target_node(void) 2372 { 2373 static int last_khugepaged_target_node = NUMA_NO_NODE; 2374 int nid, target_node = 0, max_value = 0; 2375 2376 /* find first node with max normal pages hit */ 2377 for (nid = 0; nid < MAX_NUMNODES; nid++) 2378 if (khugepaged_node_load[nid] > max_value) { 2379 max_value = khugepaged_node_load[nid]; 2380 target_node = nid; 2381 } 2382 2383 /* do some balance if several nodes have the same hit record */ 2384 if (target_node <= last_khugepaged_target_node) 2385 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES; 2386 nid++) 2387 if (max_value == khugepaged_node_load[nid]) { 2388 target_node = nid; 2389 break; 2390 } 2391 2392 last_khugepaged_target_node = target_node; 2393 return target_node; 2394 } 2395 2396 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) 2397 { 2398 if (IS_ERR(*hpage)) { 2399 if (!*wait) 2400 return false; 2401 2402 *wait = false; 2403 *hpage = NULL; 2404 khugepaged_alloc_sleep(); 2405 } else if (*hpage) { 2406 put_page(*hpage); 2407 *hpage = NULL; 2408 } 2409 2410 return true; 2411 } 2412 2413 static struct page * 2414 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm, 2415 unsigned long address, int node) 2416 { 2417 VM_BUG_ON_PAGE(*hpage, *hpage); 2418 2419 /* 2420 * Before allocating the hugepage, release the mmap_sem read lock. 2421 * The allocation can take potentially a long time if it involves 2422 * sync compaction, and we do not need to hold the mmap_sem during 2423 * that. We will recheck the vma after taking it again in write mode. 2424 */ 2425 up_read(&mm->mmap_sem); 2426 2427 *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER); 2428 if (unlikely(!*hpage)) { 2429 count_vm_event(THP_COLLAPSE_ALLOC_FAILED); 2430 *hpage = ERR_PTR(-ENOMEM); 2431 return NULL; 2432 } 2433 2434 count_vm_event(THP_COLLAPSE_ALLOC); 2435 return *hpage; 2436 } 2437 #else 2438 static int khugepaged_find_target_node(void) 2439 { 2440 return 0; 2441 } 2442 2443 static inline struct page *alloc_hugepage(int defrag) 2444 { 2445 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0), 2446 HPAGE_PMD_ORDER); 2447 } 2448 2449 static struct page *khugepaged_alloc_hugepage(bool *wait) 2450 { 2451 struct page *hpage; 2452 2453 do { 2454 hpage = alloc_hugepage(khugepaged_defrag()); 2455 if (!hpage) { 2456 count_vm_event(THP_COLLAPSE_ALLOC_FAILED); 2457 if (!*wait) 2458 return NULL; 2459 2460 *wait = false; 2461 khugepaged_alloc_sleep(); 2462 } else 2463 count_vm_event(THP_COLLAPSE_ALLOC); 2464 } while (unlikely(!hpage) && likely(khugepaged_enabled())); 2465 2466 return hpage; 2467 } 2468 2469 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) 2470 { 2471 if (!*hpage) 2472 *hpage = khugepaged_alloc_hugepage(wait); 2473 2474 if (unlikely(!*hpage)) 2475 return false; 2476 2477 return true; 2478 } 2479 2480 static struct page * 2481 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm, 2482 unsigned long address, int node) 2483 { 2484 up_read(&mm->mmap_sem); 2485 VM_BUG_ON(!*hpage); 2486 2487 return *hpage; 2488 } 2489 #endif 2490 2491 static bool hugepage_vma_check(struct vm_area_struct *vma) 2492 { 2493 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) || 2494 (vma->vm_flags & VM_NOHUGEPAGE)) 2495 return false; 2496 2497 if (!vma->anon_vma || vma->vm_ops) 2498 return false; 2499 if (is_vma_temporary_stack(vma)) 2500 return false; 2501 VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma); 2502 return true; 2503 } 2504 2505 static void collapse_huge_page(struct mm_struct *mm, 2506 unsigned long address, 2507 struct page **hpage, 2508 struct vm_area_struct *vma, 2509 int node) 2510 { 2511 pmd_t *pmd, _pmd; 2512 pte_t *pte; 2513 pgtable_t pgtable; 2514 struct page *new_page; 2515 spinlock_t *pmd_ptl, *pte_ptl; 2516 int isolated; 2517 unsigned long hstart, hend; 2518 struct mem_cgroup *memcg; 2519 unsigned long mmun_start; /* For mmu_notifiers */ 2520 unsigned long mmun_end; /* For mmu_notifiers */ 2521 gfp_t gfp; 2522 2523 VM_BUG_ON(address & ~HPAGE_PMD_MASK); 2524 2525 /* Only allocate from the target node */ 2526 gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) | 2527 __GFP_THISNODE; 2528 2529 /* release the mmap_sem read lock. */ 2530 new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node); 2531 if (!new_page) 2532 return; 2533 2534 if (unlikely(mem_cgroup_try_charge(new_page, mm, 2535 gfp, &memcg))) 2536 return; 2537 2538 /* 2539 * Prevent all access to pagetables with the exception of 2540 * gup_fast later hanlded by the ptep_clear_flush and the VM 2541 * handled by the anon_vma lock + PG_lock. 2542 */ 2543 down_write(&mm->mmap_sem); 2544 if (unlikely(khugepaged_test_exit(mm))) 2545 goto out; 2546 2547 vma = find_vma(mm, address); 2548 if (!vma) 2549 goto out; 2550 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; 2551 hend = vma->vm_end & HPAGE_PMD_MASK; 2552 if (address < hstart || address + HPAGE_PMD_SIZE > hend) 2553 goto out; 2554 if (!hugepage_vma_check(vma)) 2555 goto out; 2556 pmd = mm_find_pmd(mm, address); 2557 if (!pmd) 2558 goto out; 2559 2560 anon_vma_lock_write(vma->anon_vma); 2561 2562 pte = pte_offset_map(pmd, address); 2563 pte_ptl = pte_lockptr(mm, pmd); 2564 2565 mmun_start = address; 2566 mmun_end = address + HPAGE_PMD_SIZE; 2567 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 2568 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */ 2569 /* 2570 * After this gup_fast can't run anymore. This also removes 2571 * any huge TLB entry from the CPU so we won't allow 2572 * huge and small TLB entries for the same virtual address 2573 * to avoid the risk of CPU bugs in that area. 2574 */ 2575 _pmd = pmdp_collapse_flush(vma, address, pmd); 2576 spin_unlock(pmd_ptl); 2577 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 2578 2579 spin_lock(pte_ptl); 2580 isolated = __collapse_huge_page_isolate(vma, address, pte); 2581 spin_unlock(pte_ptl); 2582 2583 if (unlikely(!isolated)) { 2584 pte_unmap(pte); 2585 spin_lock(pmd_ptl); 2586 BUG_ON(!pmd_none(*pmd)); 2587 /* 2588 * We can only use set_pmd_at when establishing 2589 * hugepmds and never for establishing regular pmds that 2590 * points to regular pagetables. Use pmd_populate for that 2591 */ 2592 pmd_populate(mm, pmd, pmd_pgtable(_pmd)); 2593 spin_unlock(pmd_ptl); 2594 anon_vma_unlock_write(vma->anon_vma); 2595 goto out; 2596 } 2597 2598 /* 2599 * All pages are isolated and locked so anon_vma rmap 2600 * can't run anymore. 2601 */ 2602 anon_vma_unlock_write(vma->anon_vma); 2603 2604 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl); 2605 pte_unmap(pte); 2606 __SetPageUptodate(new_page); 2607 pgtable = pmd_pgtable(_pmd); 2608 2609 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot); 2610 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma); 2611 2612 /* 2613 * spin_lock() below is not the equivalent of smp_wmb(), so 2614 * this is needed to avoid the copy_huge_page writes to become 2615 * visible after the set_pmd_at() write. 2616 */ 2617 smp_wmb(); 2618 2619 spin_lock(pmd_ptl); 2620 BUG_ON(!pmd_none(*pmd)); 2621 page_add_new_anon_rmap(new_page, vma, address); 2622 mem_cgroup_commit_charge(new_page, memcg, false); 2623 lru_cache_add_active_or_unevictable(new_page, vma); 2624 pgtable_trans_huge_deposit(mm, pmd, pgtable); 2625 set_pmd_at(mm, address, pmd, _pmd); 2626 update_mmu_cache_pmd(vma, address, pmd); 2627 spin_unlock(pmd_ptl); 2628 2629 *hpage = NULL; 2630 2631 khugepaged_pages_collapsed++; 2632 out_up_write: 2633 up_write(&mm->mmap_sem); 2634 return; 2635 2636 out: 2637 mem_cgroup_cancel_charge(new_page, memcg); 2638 goto out_up_write; 2639 } 2640 2641 static int khugepaged_scan_pmd(struct mm_struct *mm, 2642 struct vm_area_struct *vma, 2643 unsigned long address, 2644 struct page **hpage) 2645 { 2646 pmd_t *pmd; 2647 pte_t *pte, *_pte; 2648 int ret = 0, none_or_zero = 0; 2649 struct page *page; 2650 unsigned long _address; 2651 spinlock_t *ptl; 2652 int node = NUMA_NO_NODE; 2653 bool writable = false, referenced = false; 2654 2655 VM_BUG_ON(address & ~HPAGE_PMD_MASK); 2656 2657 pmd = mm_find_pmd(mm, address); 2658 if (!pmd) 2659 goto out; 2660 2661 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load)); 2662 pte = pte_offset_map_lock(mm, pmd, address, &ptl); 2663 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR; 2664 _pte++, _address += PAGE_SIZE) { 2665 pte_t pteval = *_pte; 2666 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { 2667 if (!userfaultfd_armed(vma) && 2668 ++none_or_zero <= khugepaged_max_ptes_none) 2669 continue; 2670 else 2671 goto out_unmap; 2672 } 2673 if (!pte_present(pteval)) 2674 goto out_unmap; 2675 if (pte_write(pteval)) 2676 writable = true; 2677 2678 page = vm_normal_page(vma, _address, pteval); 2679 if (unlikely(!page)) 2680 goto out_unmap; 2681 /* 2682 * Record which node the original page is from and save this 2683 * information to khugepaged_node_load[]. 2684 * Khupaged will allocate hugepage from the node has the max 2685 * hit record. 2686 */ 2687 node = page_to_nid(page); 2688 if (khugepaged_scan_abort(node)) 2689 goto out_unmap; 2690 khugepaged_node_load[node]++; 2691 VM_BUG_ON_PAGE(PageCompound(page), page); 2692 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page)) 2693 goto out_unmap; 2694 /* 2695 * cannot use mapcount: can't collapse if there's a gup pin. 2696 * The page must only be referenced by the scanned process 2697 * and page swap cache. 2698 */ 2699 if (page_count(page) != 1 + !!PageSwapCache(page)) 2700 goto out_unmap; 2701 if (pte_young(pteval) || 2702 page_is_young(page) || PageReferenced(page) || 2703 mmu_notifier_test_young(vma->vm_mm, address)) 2704 referenced = true; 2705 } 2706 if (referenced && writable) 2707 ret = 1; 2708 out_unmap: 2709 pte_unmap_unlock(pte, ptl); 2710 if (ret) { 2711 node = khugepaged_find_target_node(); 2712 /* collapse_huge_page will return with the mmap_sem released */ 2713 collapse_huge_page(mm, address, hpage, vma, node); 2714 } 2715 out: 2716 return ret; 2717 } 2718 2719 static void collect_mm_slot(struct mm_slot *mm_slot) 2720 { 2721 struct mm_struct *mm = mm_slot->mm; 2722 2723 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock)); 2724 2725 if (khugepaged_test_exit(mm)) { 2726 /* free mm_slot */ 2727 hash_del(&mm_slot->hash); 2728 list_del(&mm_slot->mm_node); 2729 2730 /* 2731 * Not strictly needed because the mm exited already. 2732 * 2733 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags); 2734 */ 2735 2736 /* khugepaged_mm_lock actually not necessary for the below */ 2737 free_mm_slot(mm_slot); 2738 mmdrop(mm); 2739 } 2740 } 2741 2742 static unsigned int khugepaged_scan_mm_slot(unsigned int pages, 2743 struct page **hpage) 2744 __releases(&khugepaged_mm_lock) 2745 __acquires(&khugepaged_mm_lock) 2746 { 2747 struct mm_slot *mm_slot; 2748 struct mm_struct *mm; 2749 struct vm_area_struct *vma; 2750 int progress = 0; 2751 2752 VM_BUG_ON(!pages); 2753 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock)); 2754 2755 if (khugepaged_scan.mm_slot) 2756 mm_slot = khugepaged_scan.mm_slot; 2757 else { 2758 mm_slot = list_entry(khugepaged_scan.mm_head.next, 2759 struct mm_slot, mm_node); 2760 khugepaged_scan.address = 0; 2761 khugepaged_scan.mm_slot = mm_slot; 2762 } 2763 spin_unlock(&khugepaged_mm_lock); 2764 2765 mm = mm_slot->mm; 2766 down_read(&mm->mmap_sem); 2767 if (unlikely(khugepaged_test_exit(mm))) 2768 vma = NULL; 2769 else 2770 vma = find_vma(mm, khugepaged_scan.address); 2771 2772 progress++; 2773 for (; vma; vma = vma->vm_next) { 2774 unsigned long hstart, hend; 2775 2776 cond_resched(); 2777 if (unlikely(khugepaged_test_exit(mm))) { 2778 progress++; 2779 break; 2780 } 2781 if (!hugepage_vma_check(vma)) { 2782 skip: 2783 progress++; 2784 continue; 2785 } 2786 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; 2787 hend = vma->vm_end & HPAGE_PMD_MASK; 2788 if (hstart >= hend) 2789 goto skip; 2790 if (khugepaged_scan.address > hend) 2791 goto skip; 2792 if (khugepaged_scan.address < hstart) 2793 khugepaged_scan.address = hstart; 2794 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK); 2795 2796 while (khugepaged_scan.address < hend) { 2797 int ret; 2798 cond_resched(); 2799 if (unlikely(khugepaged_test_exit(mm))) 2800 goto breakouterloop; 2801 2802 VM_BUG_ON(khugepaged_scan.address < hstart || 2803 khugepaged_scan.address + HPAGE_PMD_SIZE > 2804 hend); 2805 ret = khugepaged_scan_pmd(mm, vma, 2806 khugepaged_scan.address, 2807 hpage); 2808 /* move to next address */ 2809 khugepaged_scan.address += HPAGE_PMD_SIZE; 2810 progress += HPAGE_PMD_NR; 2811 if (ret) 2812 /* we released mmap_sem so break loop */ 2813 goto breakouterloop_mmap_sem; 2814 if (progress >= pages) 2815 goto breakouterloop; 2816 } 2817 } 2818 breakouterloop: 2819 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */ 2820 breakouterloop_mmap_sem: 2821 2822 spin_lock(&khugepaged_mm_lock); 2823 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot); 2824 /* 2825 * Release the current mm_slot if this mm is about to die, or 2826 * if we scanned all vmas of this mm. 2827 */ 2828 if (khugepaged_test_exit(mm) || !vma) { 2829 /* 2830 * Make sure that if mm_users is reaching zero while 2831 * khugepaged runs here, khugepaged_exit will find 2832 * mm_slot not pointing to the exiting mm. 2833 */ 2834 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) { 2835 khugepaged_scan.mm_slot = list_entry( 2836 mm_slot->mm_node.next, 2837 struct mm_slot, mm_node); 2838 khugepaged_scan.address = 0; 2839 } else { 2840 khugepaged_scan.mm_slot = NULL; 2841 khugepaged_full_scans++; 2842 } 2843 2844 collect_mm_slot(mm_slot); 2845 } 2846 2847 return progress; 2848 } 2849 2850 static int khugepaged_has_work(void) 2851 { 2852 return !list_empty(&khugepaged_scan.mm_head) && 2853 khugepaged_enabled(); 2854 } 2855 2856 static int khugepaged_wait_event(void) 2857 { 2858 return !list_empty(&khugepaged_scan.mm_head) || 2859 kthread_should_stop(); 2860 } 2861 2862 static void khugepaged_do_scan(void) 2863 { 2864 struct page *hpage = NULL; 2865 unsigned int progress = 0, pass_through_head = 0; 2866 unsigned int pages = khugepaged_pages_to_scan; 2867 bool wait = true; 2868 2869 barrier(); /* write khugepaged_pages_to_scan to local stack */ 2870 2871 while (progress < pages) { 2872 if (!khugepaged_prealloc_page(&hpage, &wait)) 2873 break; 2874 2875 cond_resched(); 2876 2877 if (unlikely(kthread_should_stop() || try_to_freeze())) 2878 break; 2879 2880 spin_lock(&khugepaged_mm_lock); 2881 if (!khugepaged_scan.mm_slot) 2882 pass_through_head++; 2883 if (khugepaged_has_work() && 2884 pass_through_head < 2) 2885 progress += khugepaged_scan_mm_slot(pages - progress, 2886 &hpage); 2887 else 2888 progress = pages; 2889 spin_unlock(&khugepaged_mm_lock); 2890 } 2891 2892 if (!IS_ERR_OR_NULL(hpage)) 2893 put_page(hpage); 2894 } 2895 2896 static void khugepaged_wait_work(void) 2897 { 2898 if (khugepaged_has_work()) { 2899 if (!khugepaged_scan_sleep_millisecs) 2900 return; 2901 2902 wait_event_freezable_timeout(khugepaged_wait, 2903 kthread_should_stop(), 2904 msecs_to_jiffies(khugepaged_scan_sleep_millisecs)); 2905 return; 2906 } 2907 2908 if (khugepaged_enabled()) 2909 wait_event_freezable(khugepaged_wait, khugepaged_wait_event()); 2910 } 2911 2912 static int khugepaged(void *none) 2913 { 2914 struct mm_slot *mm_slot; 2915 2916 set_freezable(); 2917 set_user_nice(current, MAX_NICE); 2918 2919 while (!kthread_should_stop()) { 2920 khugepaged_do_scan(); 2921 khugepaged_wait_work(); 2922 } 2923 2924 spin_lock(&khugepaged_mm_lock); 2925 mm_slot = khugepaged_scan.mm_slot; 2926 khugepaged_scan.mm_slot = NULL; 2927 if (mm_slot) 2928 collect_mm_slot(mm_slot); 2929 spin_unlock(&khugepaged_mm_lock); 2930 return 0; 2931 } 2932 2933 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, 2934 unsigned long haddr, pmd_t *pmd) 2935 { 2936 struct mm_struct *mm = vma->vm_mm; 2937 pgtable_t pgtable; 2938 pmd_t _pmd; 2939 int i; 2940 2941 pmdp_huge_clear_flush_notify(vma, haddr, pmd); 2942 /* leave pmd empty until pte is filled */ 2943 2944 pgtable = pgtable_trans_huge_withdraw(mm, pmd); 2945 pmd_populate(mm, &_pmd, pgtable); 2946 2947 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { 2948 pte_t *pte, entry; 2949 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot); 2950 entry = pte_mkspecial(entry); 2951 pte = pte_offset_map(&_pmd, haddr); 2952 VM_BUG_ON(!pte_none(*pte)); 2953 set_pte_at(mm, haddr, pte, entry); 2954 pte_unmap(pte); 2955 } 2956 smp_wmb(); /* make pte visible before pmd */ 2957 pmd_populate(mm, pmd, pgtable); 2958 put_huge_zero_page(); 2959 } 2960 2961 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address, 2962 pmd_t *pmd) 2963 { 2964 spinlock_t *ptl; 2965 struct page *page = NULL; 2966 struct mm_struct *mm = vma->vm_mm; 2967 unsigned long haddr = address & HPAGE_PMD_MASK; 2968 unsigned long mmun_start; /* For mmu_notifiers */ 2969 unsigned long mmun_end; /* For mmu_notifiers */ 2970 2971 BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE); 2972 2973 mmun_start = haddr; 2974 mmun_end = haddr + HPAGE_PMD_SIZE; 2975 again: 2976 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 2977 ptl = pmd_lock(mm, pmd); 2978 if (unlikely(!pmd_trans_huge(*pmd))) 2979 goto unlock; 2980 if (vma_is_dax(vma)) { 2981 pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd); 2982 if (is_huge_zero_pmd(_pmd)) 2983 put_huge_zero_page(); 2984 } else if (is_huge_zero_pmd(*pmd)) { 2985 __split_huge_zero_page_pmd(vma, haddr, pmd); 2986 } else { 2987 page = pmd_page(*pmd); 2988 VM_BUG_ON_PAGE(!page_count(page), page); 2989 get_page(page); 2990 } 2991 unlock: 2992 spin_unlock(ptl); 2993 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 2994 2995 if (!page) 2996 return; 2997 2998 split_huge_page(page); 2999 put_page(page); 3000 3001 /* 3002 * We don't always have down_write of mmap_sem here: a racing 3003 * do_huge_pmd_wp_page() might have copied-on-write to another 3004 * huge page before our split_huge_page() got the anon_vma lock. 3005 */ 3006 if (unlikely(pmd_trans_huge(*pmd))) 3007 goto again; 3008 } 3009 3010 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address, 3011 pmd_t *pmd) 3012 { 3013 struct vm_area_struct *vma; 3014 3015 vma = find_vma(mm, address); 3016 BUG_ON(vma == NULL); 3017 split_huge_page_pmd(vma, address, pmd); 3018 } 3019 3020 static void split_huge_page_address(struct mm_struct *mm, 3021 unsigned long address) 3022 { 3023 pgd_t *pgd; 3024 pud_t *pud; 3025 pmd_t *pmd; 3026 3027 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK)); 3028 3029 pgd = pgd_offset(mm, address); 3030 if (!pgd_present(*pgd)) 3031 return; 3032 3033 pud = pud_offset(pgd, address); 3034 if (!pud_present(*pud)) 3035 return; 3036 3037 pmd = pmd_offset(pud, address); 3038 if (!pmd_present(*pmd)) 3039 return; 3040 /* 3041 * Caller holds the mmap_sem write mode, so a huge pmd cannot 3042 * materialize from under us. 3043 */ 3044 split_huge_page_pmd_mm(mm, address, pmd); 3045 } 3046 3047 void vma_adjust_trans_huge(struct vm_area_struct *vma, 3048 unsigned long start, 3049 unsigned long end, 3050 long adjust_next) 3051 { 3052 /* 3053 * If the new start address isn't hpage aligned and it could 3054 * previously contain an hugepage: check if we need to split 3055 * an huge pmd. 3056 */ 3057 if (start & ~HPAGE_PMD_MASK && 3058 (start & HPAGE_PMD_MASK) >= vma->vm_start && 3059 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) 3060 split_huge_page_address(vma->vm_mm, start); 3061 3062 /* 3063 * If the new end address isn't hpage aligned and it could 3064 * previously contain an hugepage: check if we need to split 3065 * an huge pmd. 3066 */ 3067 if (end & ~HPAGE_PMD_MASK && 3068 (end & HPAGE_PMD_MASK) >= vma->vm_start && 3069 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) 3070 split_huge_page_address(vma->vm_mm, end); 3071 3072 /* 3073 * If we're also updating the vma->vm_next->vm_start, if the new 3074 * vm_next->vm_start isn't page aligned and it could previously 3075 * contain an hugepage: check if we need to split an huge pmd. 3076 */ 3077 if (adjust_next > 0) { 3078 struct vm_area_struct *next = vma->vm_next; 3079 unsigned long nstart = next->vm_start; 3080 nstart += adjust_next << PAGE_SHIFT; 3081 if (nstart & ~HPAGE_PMD_MASK && 3082 (nstart & HPAGE_PMD_MASK) >= next->vm_start && 3083 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end) 3084 split_huge_page_address(next->vm_mm, nstart); 3085 } 3086 } 3087