ksm.c revision 08beca44dfb0ab008e365163df70dbd302ae1508
1/* 2 * Memory merging support. 3 * 4 * This code enables dynamic sharing of identical pages found in different 5 * memory areas, even if they are not shared by fork() 6 * 7 * Copyright (C) 2008-2009 Red Hat, Inc. 8 * Authors: 9 * Izik Eidus 10 * Andrea Arcangeli 11 * Chris Wright 12 * Hugh Dickins 13 * 14 * This work is licensed under the terms of the GNU GPL, version 2. 15 */ 16 17#include <linux/errno.h> 18#include <linux/mm.h> 19#include <linux/fs.h> 20#include <linux/mman.h> 21#include <linux/sched.h> 22#include <linux/rwsem.h> 23#include <linux/pagemap.h> 24#include <linux/rmap.h> 25#include <linux/spinlock.h> 26#include <linux/jhash.h> 27#include <linux/delay.h> 28#include <linux/kthread.h> 29#include <linux/wait.h> 30#include <linux/slab.h> 31#include <linux/rbtree.h> 32#include <linux/mmu_notifier.h> 33#include <linux/swap.h> 34#include <linux/ksm.h> 35 36#include <asm/tlbflush.h> 37 38/* 39 * A few notes about the KSM scanning process, 40 * to make it easier to understand the data structures below: 41 * 42 * In order to reduce excessive scanning, KSM sorts the memory pages by their 43 * contents into a data structure that holds pointers to the pages' locations. 44 * 45 * Since the contents of the pages may change at any moment, KSM cannot just 46 * insert the pages into a normal sorted tree and expect it to find anything. 47 * Therefore KSM uses two data structures - the stable and the unstable tree. 48 * 49 * The stable tree holds pointers to all the merged pages (ksm pages), sorted 50 * by their contents. Because each such page is write-protected, searching on 51 * this tree is fully assured to be working (except when pages are unmapped), 52 * and therefore this tree is called the stable tree. 53 * 54 * In addition to the stable tree, KSM uses a second data structure called the 55 * unstable tree: this tree holds pointers to pages which have been found to 56 * be "unchanged for a period of time". The unstable tree sorts these pages 57 * by their contents, but since they are not write-protected, KSM cannot rely 58 * upon the unstable tree to work correctly - the unstable tree is liable to 59 * be corrupted as its contents are modified, and so it is called unstable. 60 * 61 * KSM solves this problem by several techniques: 62 * 63 * 1) The unstable tree is flushed every time KSM completes scanning all 64 * memory areas, and then the tree is rebuilt again from the beginning. 65 * 2) KSM will only insert into the unstable tree, pages whose hash value 66 * has not changed since the previous scan of all memory areas. 67 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the 68 * colors of the nodes and not on their contents, assuring that even when 69 * the tree gets "corrupted" it won't get out of balance, so scanning time 70 * remains the same (also, searching and inserting nodes in an rbtree uses 71 * the same algorithm, so we have no overhead when we flush and rebuild). 72 * 4) KSM never flushes the stable tree, which means that even if it were to 73 * take 10 attempts to find a page in the unstable tree, once it is found, 74 * it is secured in the stable tree. (When we scan a new page, we first 75 * compare it against the stable tree, and then against the unstable tree.) 76 */ 77 78/** 79 * struct mm_slot - ksm information per mm that is being scanned 80 * @link: link to the mm_slots hash list 81 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head 82 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items 83 * @mm: the mm that this information is valid for 84 */ 85struct mm_slot { 86 struct hlist_node link; 87 struct list_head mm_list; 88 struct rmap_item *rmap_list; 89 struct mm_struct *mm; 90}; 91 92/** 93 * struct ksm_scan - cursor for scanning 94 * @mm_slot: the current mm_slot we are scanning 95 * @address: the next address inside that to be scanned 96 * @rmap_list: link to the next rmap to be scanned in the rmap_list 97 * @seqnr: count of completed full scans (needed when removing unstable node) 98 * 99 * There is only the one ksm_scan instance of this cursor structure. 100 */ 101struct ksm_scan { 102 struct mm_slot *mm_slot; 103 unsigned long address; 104 struct rmap_item **rmap_list; 105 unsigned long seqnr; 106}; 107 108/** 109 * struct stable_node - node of the stable rbtree 110 * @page: pointer to struct page of the ksm page 111 * @node: rb node of this ksm page in the stable tree 112 * @hlist: hlist head of rmap_items using this ksm page 113 */ 114struct stable_node { 115 struct page *page; 116 struct rb_node node; 117 struct hlist_head hlist; 118}; 119 120/** 121 * struct rmap_item - reverse mapping item for virtual addresses 122 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list 123 * @filler: unused space we're making available in this patch 124 * @mm: the memory structure this rmap_item is pointing into 125 * @address: the virtual address this rmap_item tracks (+ flags in low bits) 126 * @oldchecksum: previous checksum of the page at that virtual address 127 * @node: rb node of this rmap_item in the unstable tree 128 * @head: pointer to stable_node heading this list in the stable tree 129 * @hlist: link into hlist of rmap_items hanging off that stable_node 130 */ 131struct rmap_item { 132 struct rmap_item *rmap_list; 133 unsigned long filler; 134 struct mm_struct *mm; 135 unsigned long address; /* + low bits used for flags below */ 136 unsigned int oldchecksum; /* when unstable */ 137 union { 138 struct rb_node node; /* when node of unstable tree */ 139 struct { /* when listed from stable tree */ 140 struct stable_node *head; 141 struct hlist_node hlist; 142 }; 143 }; 144}; 145 146#define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */ 147#define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */ 148#define STABLE_FLAG 0x200 /* is listed from the stable tree */ 149 150/* The stable and unstable tree heads */ 151static struct rb_root root_stable_tree = RB_ROOT; 152static struct rb_root root_unstable_tree = RB_ROOT; 153 154#define MM_SLOTS_HASH_HEADS 1024 155static struct hlist_head *mm_slots_hash; 156 157static struct mm_slot ksm_mm_head = { 158 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list), 159}; 160static struct ksm_scan ksm_scan = { 161 .mm_slot = &ksm_mm_head, 162}; 163 164static struct kmem_cache *rmap_item_cache; 165static struct kmem_cache *stable_node_cache; 166static struct kmem_cache *mm_slot_cache; 167 168/* The number of nodes in the stable tree */ 169static unsigned long ksm_pages_shared; 170 171/* The number of page slots additionally sharing those nodes */ 172static unsigned long ksm_pages_sharing; 173 174/* The number of nodes in the unstable tree */ 175static unsigned long ksm_pages_unshared; 176 177/* The number of rmap_items in use: to calculate pages_volatile */ 178static unsigned long ksm_rmap_items; 179 180/* Limit on the number of unswappable pages used */ 181static unsigned long ksm_max_kernel_pages; 182 183/* Number of pages ksmd should scan in one batch */ 184static unsigned int ksm_thread_pages_to_scan = 100; 185 186/* Milliseconds ksmd should sleep between batches */ 187static unsigned int ksm_thread_sleep_millisecs = 20; 188 189#define KSM_RUN_STOP 0 190#define KSM_RUN_MERGE 1 191#define KSM_RUN_UNMERGE 2 192static unsigned int ksm_run = KSM_RUN_STOP; 193 194static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait); 195static DEFINE_MUTEX(ksm_thread_mutex); 196static DEFINE_SPINLOCK(ksm_mmlist_lock); 197 198#define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\ 199 sizeof(struct __struct), __alignof__(struct __struct),\ 200 (__flags), NULL) 201 202static int __init ksm_slab_init(void) 203{ 204 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0); 205 if (!rmap_item_cache) 206 goto out; 207 208 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0); 209 if (!stable_node_cache) 210 goto out_free1; 211 212 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0); 213 if (!mm_slot_cache) 214 goto out_free2; 215 216 return 0; 217 218out_free2: 219 kmem_cache_destroy(stable_node_cache); 220out_free1: 221 kmem_cache_destroy(rmap_item_cache); 222out: 223 return -ENOMEM; 224} 225 226static void __init ksm_slab_free(void) 227{ 228 kmem_cache_destroy(mm_slot_cache); 229 kmem_cache_destroy(stable_node_cache); 230 kmem_cache_destroy(rmap_item_cache); 231 mm_slot_cache = NULL; 232} 233 234static inline struct rmap_item *alloc_rmap_item(void) 235{ 236 struct rmap_item *rmap_item; 237 238 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL); 239 if (rmap_item) 240 ksm_rmap_items++; 241 return rmap_item; 242} 243 244static inline void free_rmap_item(struct rmap_item *rmap_item) 245{ 246 ksm_rmap_items--; 247 rmap_item->mm = NULL; /* debug safety */ 248 kmem_cache_free(rmap_item_cache, rmap_item); 249} 250 251static inline struct stable_node *alloc_stable_node(void) 252{ 253 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL); 254} 255 256static inline void free_stable_node(struct stable_node *stable_node) 257{ 258 kmem_cache_free(stable_node_cache, stable_node); 259} 260 261static inline struct mm_slot *alloc_mm_slot(void) 262{ 263 if (!mm_slot_cache) /* initialization failed */ 264 return NULL; 265 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL); 266} 267 268static inline void free_mm_slot(struct mm_slot *mm_slot) 269{ 270 kmem_cache_free(mm_slot_cache, mm_slot); 271} 272 273static int __init mm_slots_hash_init(void) 274{ 275 mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head), 276 GFP_KERNEL); 277 if (!mm_slots_hash) 278 return -ENOMEM; 279 return 0; 280} 281 282static void __init mm_slots_hash_free(void) 283{ 284 kfree(mm_slots_hash); 285} 286 287static struct mm_slot *get_mm_slot(struct mm_struct *mm) 288{ 289 struct mm_slot *mm_slot; 290 struct hlist_head *bucket; 291 struct hlist_node *node; 292 293 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct)) 294 % MM_SLOTS_HASH_HEADS]; 295 hlist_for_each_entry(mm_slot, node, bucket, link) { 296 if (mm == mm_slot->mm) 297 return mm_slot; 298 } 299 return NULL; 300} 301 302static void insert_to_mm_slots_hash(struct mm_struct *mm, 303 struct mm_slot *mm_slot) 304{ 305 struct hlist_head *bucket; 306 307 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct)) 308 % MM_SLOTS_HASH_HEADS]; 309 mm_slot->mm = mm; 310 hlist_add_head(&mm_slot->link, bucket); 311} 312 313static inline int in_stable_tree(struct rmap_item *rmap_item) 314{ 315 return rmap_item->address & STABLE_FLAG; 316} 317 318/* 319 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's 320 * page tables after it has passed through ksm_exit() - which, if necessary, 321 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set 322 * a special flag: they can just back out as soon as mm_users goes to zero. 323 * ksm_test_exit() is used throughout to make this test for exit: in some 324 * places for correctness, in some places just to avoid unnecessary work. 325 */ 326static inline bool ksm_test_exit(struct mm_struct *mm) 327{ 328 return atomic_read(&mm->mm_users) == 0; 329} 330 331/* 332 * We use break_ksm to break COW on a ksm page: it's a stripped down 333 * 334 * if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1) 335 * put_page(page); 336 * 337 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma, 338 * in case the application has unmapped and remapped mm,addr meanwhile. 339 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP 340 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it. 341 */ 342static int break_ksm(struct vm_area_struct *vma, unsigned long addr) 343{ 344 struct page *page; 345 int ret = 0; 346 347 do { 348 cond_resched(); 349 page = follow_page(vma, addr, FOLL_GET); 350 if (!page) 351 break; 352 if (PageKsm(page)) 353 ret = handle_mm_fault(vma->vm_mm, vma, addr, 354 FAULT_FLAG_WRITE); 355 else 356 ret = VM_FAULT_WRITE; 357 put_page(page); 358 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM))); 359 /* 360 * We must loop because handle_mm_fault() may back out if there's 361 * any difficulty e.g. if pte accessed bit gets updated concurrently. 362 * 363 * VM_FAULT_WRITE is what we have been hoping for: it indicates that 364 * COW has been broken, even if the vma does not permit VM_WRITE; 365 * but note that a concurrent fault might break PageKsm for us. 366 * 367 * VM_FAULT_SIGBUS could occur if we race with truncation of the 368 * backing file, which also invalidates anonymous pages: that's 369 * okay, that truncation will have unmapped the PageKsm for us. 370 * 371 * VM_FAULT_OOM: at the time of writing (late July 2009), setting 372 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the 373 * current task has TIF_MEMDIE set, and will be OOM killed on return 374 * to user; and ksmd, having no mm, would never be chosen for that. 375 * 376 * But if the mm is in a limited mem_cgroup, then the fault may fail 377 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and 378 * even ksmd can fail in this way - though it's usually breaking ksm 379 * just to undo a merge it made a moment before, so unlikely to oom. 380 * 381 * That's a pity: we might therefore have more kernel pages allocated 382 * than we're counting as nodes in the stable tree; but ksm_do_scan 383 * will retry to break_cow on each pass, so should recover the page 384 * in due course. The important thing is to not let VM_MERGEABLE 385 * be cleared while any such pages might remain in the area. 386 */ 387 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0; 388} 389 390static void break_cow(struct rmap_item *rmap_item) 391{ 392 struct mm_struct *mm = rmap_item->mm; 393 unsigned long addr = rmap_item->address; 394 struct vm_area_struct *vma; 395 396 down_read(&mm->mmap_sem); 397 if (ksm_test_exit(mm)) 398 goto out; 399 vma = find_vma(mm, addr); 400 if (!vma || vma->vm_start > addr) 401 goto out; 402 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) 403 goto out; 404 break_ksm(vma, addr); 405out: 406 up_read(&mm->mmap_sem); 407} 408 409static struct page *get_mergeable_page(struct rmap_item *rmap_item) 410{ 411 struct mm_struct *mm = rmap_item->mm; 412 unsigned long addr = rmap_item->address; 413 struct vm_area_struct *vma; 414 struct page *page; 415 416 down_read(&mm->mmap_sem); 417 if (ksm_test_exit(mm)) 418 goto out; 419 vma = find_vma(mm, addr); 420 if (!vma || vma->vm_start > addr) 421 goto out; 422 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) 423 goto out; 424 425 page = follow_page(vma, addr, FOLL_GET); 426 if (!page) 427 goto out; 428 if (PageAnon(page)) { 429 flush_anon_page(vma, page, addr); 430 flush_dcache_page(page); 431 } else { 432 put_page(page); 433out: page = NULL; 434 } 435 up_read(&mm->mmap_sem); 436 return page; 437} 438 439/* 440 * Removing rmap_item from stable or unstable tree. 441 * This function will clean the information from the stable/unstable tree. 442 */ 443static void remove_rmap_item_from_tree(struct rmap_item *rmap_item) 444{ 445 if (rmap_item->address & STABLE_FLAG) { 446 struct stable_node *stable_node; 447 448 stable_node = rmap_item->head; 449 hlist_del(&rmap_item->hlist); 450 if (stable_node->hlist.first) 451 ksm_pages_sharing--; 452 else { 453 set_page_stable_node(stable_node->page, NULL); 454 put_page(stable_node->page); 455 456 rb_erase(&stable_node->node, &root_stable_tree); 457 free_stable_node(stable_node); 458 ksm_pages_shared--; 459 } 460 461 rmap_item->address &= PAGE_MASK; 462 463 } else if (rmap_item->address & UNSTABLE_FLAG) { 464 unsigned char age; 465 /* 466 * Usually ksmd can and must skip the rb_erase, because 467 * root_unstable_tree was already reset to RB_ROOT. 468 * But be careful when an mm is exiting: do the rb_erase 469 * if this rmap_item was inserted by this scan, rather 470 * than left over from before. 471 */ 472 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address); 473 BUG_ON(age > 1); 474 if (!age) 475 rb_erase(&rmap_item->node, &root_unstable_tree); 476 477 ksm_pages_unshared--; 478 rmap_item->address &= PAGE_MASK; 479 } 480 481 cond_resched(); /* we're called from many long loops */ 482} 483 484static void remove_trailing_rmap_items(struct mm_slot *mm_slot, 485 struct rmap_item **rmap_list) 486{ 487 while (*rmap_list) { 488 struct rmap_item *rmap_item = *rmap_list; 489 *rmap_list = rmap_item->rmap_list; 490 remove_rmap_item_from_tree(rmap_item); 491 free_rmap_item(rmap_item); 492 } 493} 494 495/* 496 * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather 497 * than check every pte of a given vma, the locking doesn't quite work for 498 * that - an rmap_item is assigned to the stable tree after inserting ksm 499 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing 500 * rmap_items from parent to child at fork time (so as not to waste time 501 * if exit comes before the next scan reaches it). 502 * 503 * Similarly, although we'd like to remove rmap_items (so updating counts 504 * and freeing memory) when unmerging an area, it's easier to leave that 505 * to the next pass of ksmd - consider, for example, how ksmd might be 506 * in cmp_and_merge_page on one of the rmap_items we would be removing. 507 */ 508static int unmerge_ksm_pages(struct vm_area_struct *vma, 509 unsigned long start, unsigned long end) 510{ 511 unsigned long addr; 512 int err = 0; 513 514 for (addr = start; addr < end && !err; addr += PAGE_SIZE) { 515 if (ksm_test_exit(vma->vm_mm)) 516 break; 517 if (signal_pending(current)) 518 err = -ERESTARTSYS; 519 else 520 err = break_ksm(vma, addr); 521 } 522 return err; 523} 524 525#ifdef CONFIG_SYSFS 526/* 527 * Only called through the sysfs control interface: 528 */ 529static int unmerge_and_remove_all_rmap_items(void) 530{ 531 struct mm_slot *mm_slot; 532 struct mm_struct *mm; 533 struct vm_area_struct *vma; 534 int err = 0; 535 536 spin_lock(&ksm_mmlist_lock); 537 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next, 538 struct mm_slot, mm_list); 539 spin_unlock(&ksm_mmlist_lock); 540 541 for (mm_slot = ksm_scan.mm_slot; 542 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) { 543 mm = mm_slot->mm; 544 down_read(&mm->mmap_sem); 545 for (vma = mm->mmap; vma; vma = vma->vm_next) { 546 if (ksm_test_exit(mm)) 547 break; 548 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) 549 continue; 550 err = unmerge_ksm_pages(vma, 551 vma->vm_start, vma->vm_end); 552 if (err) 553 goto error; 554 } 555 556 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list); 557 558 spin_lock(&ksm_mmlist_lock); 559 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next, 560 struct mm_slot, mm_list); 561 if (ksm_test_exit(mm)) { 562 hlist_del(&mm_slot->link); 563 list_del(&mm_slot->mm_list); 564 spin_unlock(&ksm_mmlist_lock); 565 566 free_mm_slot(mm_slot); 567 clear_bit(MMF_VM_MERGEABLE, &mm->flags); 568 up_read(&mm->mmap_sem); 569 mmdrop(mm); 570 } else { 571 spin_unlock(&ksm_mmlist_lock); 572 up_read(&mm->mmap_sem); 573 } 574 } 575 576 ksm_scan.seqnr = 0; 577 return 0; 578 579error: 580 up_read(&mm->mmap_sem); 581 spin_lock(&ksm_mmlist_lock); 582 ksm_scan.mm_slot = &ksm_mm_head; 583 spin_unlock(&ksm_mmlist_lock); 584 return err; 585} 586#endif /* CONFIG_SYSFS */ 587 588static u32 calc_checksum(struct page *page) 589{ 590 u32 checksum; 591 void *addr = kmap_atomic(page, KM_USER0); 592 checksum = jhash2(addr, PAGE_SIZE / 4, 17); 593 kunmap_atomic(addr, KM_USER0); 594 return checksum; 595} 596 597static int memcmp_pages(struct page *page1, struct page *page2) 598{ 599 char *addr1, *addr2; 600 int ret; 601 602 addr1 = kmap_atomic(page1, KM_USER0); 603 addr2 = kmap_atomic(page2, KM_USER1); 604 ret = memcmp(addr1, addr2, PAGE_SIZE); 605 kunmap_atomic(addr2, KM_USER1); 606 kunmap_atomic(addr1, KM_USER0); 607 return ret; 608} 609 610static inline int pages_identical(struct page *page1, struct page *page2) 611{ 612 return !memcmp_pages(page1, page2); 613} 614 615static int write_protect_page(struct vm_area_struct *vma, struct page *page, 616 pte_t *orig_pte) 617{ 618 struct mm_struct *mm = vma->vm_mm; 619 unsigned long addr; 620 pte_t *ptep; 621 spinlock_t *ptl; 622 int swapped; 623 int err = -EFAULT; 624 625 addr = page_address_in_vma(page, vma); 626 if (addr == -EFAULT) 627 goto out; 628 629 ptep = page_check_address(page, mm, addr, &ptl, 0); 630 if (!ptep) 631 goto out; 632 633 if (pte_write(*ptep)) { 634 pte_t entry; 635 636 swapped = PageSwapCache(page); 637 flush_cache_page(vma, addr, page_to_pfn(page)); 638 /* 639 * Ok this is tricky, when get_user_pages_fast() run it doesnt 640 * take any lock, therefore the check that we are going to make 641 * with the pagecount against the mapcount is racey and 642 * O_DIRECT can happen right after the check. 643 * So we clear the pte and flush the tlb before the check 644 * this assure us that no O_DIRECT can happen after the check 645 * or in the middle of the check. 646 */ 647 entry = ptep_clear_flush(vma, addr, ptep); 648 /* 649 * Check that no O_DIRECT or similar I/O is in progress on the 650 * page 651 */ 652 if (page_mapcount(page) + 1 + swapped != page_count(page)) { 653 set_pte_at_notify(mm, addr, ptep, entry); 654 goto out_unlock; 655 } 656 entry = pte_wrprotect(entry); 657 set_pte_at_notify(mm, addr, ptep, entry); 658 } 659 *orig_pte = *ptep; 660 err = 0; 661 662out_unlock: 663 pte_unmap_unlock(ptep, ptl); 664out: 665 return err; 666} 667 668/** 669 * replace_page - replace page in vma by new ksm page 670 * @vma: vma that holds the pte pointing to page 671 * @page: the page we are replacing by kpage 672 * @kpage: the ksm page we replace page by 673 * @orig_pte: the original value of the pte 674 * 675 * Returns 0 on success, -EFAULT on failure. 676 */ 677static int replace_page(struct vm_area_struct *vma, struct page *page, 678 struct page *kpage, pte_t orig_pte) 679{ 680 struct mm_struct *mm = vma->vm_mm; 681 pgd_t *pgd; 682 pud_t *pud; 683 pmd_t *pmd; 684 pte_t *ptep; 685 spinlock_t *ptl; 686 unsigned long addr; 687 int err = -EFAULT; 688 689 addr = page_address_in_vma(page, vma); 690 if (addr == -EFAULT) 691 goto out; 692 693 pgd = pgd_offset(mm, addr); 694 if (!pgd_present(*pgd)) 695 goto out; 696 697 pud = pud_offset(pgd, addr); 698 if (!pud_present(*pud)) 699 goto out; 700 701 pmd = pmd_offset(pud, addr); 702 if (!pmd_present(*pmd)) 703 goto out; 704 705 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl); 706 if (!pte_same(*ptep, orig_pte)) { 707 pte_unmap_unlock(ptep, ptl); 708 goto out; 709 } 710 711 get_page(kpage); 712 page_add_ksm_rmap(kpage); 713 714 flush_cache_page(vma, addr, pte_pfn(*ptep)); 715 ptep_clear_flush(vma, addr, ptep); 716 set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot)); 717 718 page_remove_rmap(page); 719 put_page(page); 720 721 pte_unmap_unlock(ptep, ptl); 722 err = 0; 723out: 724 return err; 725} 726 727/* 728 * try_to_merge_one_page - take two pages and merge them into one 729 * @vma: the vma that holds the pte pointing to page 730 * @page: the PageAnon page that we want to replace with kpage 731 * @kpage: the PageKsm page that we want to map instead of page 732 * 733 * This function returns 0 if the pages were merged, -EFAULT otherwise. 734 */ 735static int try_to_merge_one_page(struct vm_area_struct *vma, 736 struct page *page, struct page *kpage) 737{ 738 pte_t orig_pte = __pte(0); 739 int err = -EFAULT; 740 741 if (!(vma->vm_flags & VM_MERGEABLE)) 742 goto out; 743 if (!PageAnon(page)) 744 goto out; 745 746 /* 747 * We need the page lock to read a stable PageSwapCache in 748 * write_protect_page(). We use trylock_page() instead of 749 * lock_page() because we don't want to wait here - we 750 * prefer to continue scanning and merging different pages, 751 * then come back to this page when it is unlocked. 752 */ 753 if (!trylock_page(page)) 754 goto out; 755 /* 756 * If this anonymous page is mapped only here, its pte may need 757 * to be write-protected. If it's mapped elsewhere, all of its 758 * ptes are necessarily already write-protected. But in either 759 * case, we need to lock and check page_count is not raised. 760 */ 761 if (write_protect_page(vma, page, &orig_pte) == 0 && 762 pages_identical(page, kpage)) 763 err = replace_page(vma, page, kpage, orig_pte); 764 765 unlock_page(page); 766out: 767 return err; 768} 769 770/* 771 * try_to_merge_with_ksm_page - like try_to_merge_two_pages, 772 * but no new kernel page is allocated: kpage must already be a ksm page. 773 * 774 * This function returns 0 if the pages were merged, -EFAULT otherwise. 775 */ 776static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item, 777 struct page *page, struct page *kpage) 778{ 779 struct mm_struct *mm = rmap_item->mm; 780 struct vm_area_struct *vma; 781 int err = -EFAULT; 782 783 if (page == kpage) /* ksm page forked */ 784 return 0; 785 786 down_read(&mm->mmap_sem); 787 if (ksm_test_exit(mm)) 788 goto out; 789 vma = find_vma(mm, rmap_item->address); 790 if (!vma || vma->vm_start > rmap_item->address) 791 goto out; 792 793 err = try_to_merge_one_page(vma, page, kpage); 794out: 795 up_read(&mm->mmap_sem); 796 return err; 797} 798 799/* 800 * try_to_merge_two_pages - take two identical pages and prepare them 801 * to be merged into one page. 802 * 803 * This function returns the kpage if we successfully merged two identical 804 * pages into one ksm page, NULL otherwise. 805 * 806 * Note that this function allocates a new kernel page: if one of the pages 807 * is already a ksm page, try_to_merge_with_ksm_page should be used. 808 */ 809static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item, 810 struct page *page, 811 struct rmap_item *tree_rmap_item, 812 struct page *tree_page) 813{ 814 struct mm_struct *mm = rmap_item->mm; 815 struct vm_area_struct *vma; 816 struct page *kpage; 817 int err = -EFAULT; 818 819 /* 820 * The number of nodes in the stable tree 821 * is the number of kernel pages that we hold. 822 */ 823 if (ksm_max_kernel_pages && 824 ksm_max_kernel_pages <= ksm_pages_shared) 825 return NULL; 826 827 kpage = alloc_page(GFP_HIGHUSER); 828 if (!kpage) 829 return NULL; 830 831 down_read(&mm->mmap_sem); 832 if (ksm_test_exit(mm)) 833 goto up; 834 vma = find_vma(mm, rmap_item->address); 835 if (!vma || vma->vm_start > rmap_item->address) 836 goto up; 837 838 copy_user_highpage(kpage, page, rmap_item->address, vma); 839 840 set_page_stable_node(kpage, NULL); /* mark it PageKsm */ 841 842 err = try_to_merge_one_page(vma, page, kpage); 843up: 844 up_read(&mm->mmap_sem); 845 846 if (!err) { 847 err = try_to_merge_with_ksm_page(tree_rmap_item, 848 tree_page, kpage); 849 /* 850 * If that fails, we have a ksm page with only one pte 851 * pointing to it: so break it. 852 */ 853 if (err) 854 break_cow(rmap_item); 855 } 856 if (err) { 857 put_page(kpage); 858 kpage = NULL; 859 } 860 return kpage; 861} 862 863/* 864 * stable_tree_search - search for page inside the stable tree 865 * 866 * This function checks if there is a page inside the stable tree 867 * with identical content to the page that we are scanning right now. 868 * 869 * This function returns the stable tree node of identical content if found, 870 * NULL otherwise. 871 */ 872static struct stable_node *stable_tree_search(struct page *page) 873{ 874 struct rb_node *node = root_stable_tree.rb_node; 875 struct stable_node *stable_node; 876 877 stable_node = page_stable_node(page); 878 if (stable_node) { /* ksm page forked */ 879 get_page(page); 880 return stable_node; 881 } 882 883 while (node) { 884 int ret; 885 886 cond_resched(); 887 stable_node = rb_entry(node, struct stable_node, node); 888 889 ret = memcmp_pages(page, stable_node->page); 890 891 if (ret < 0) 892 node = node->rb_left; 893 else if (ret > 0) 894 node = node->rb_right; 895 else { 896 get_page(stable_node->page); 897 return stable_node; 898 } 899 } 900 901 return NULL; 902} 903 904/* 905 * stable_tree_insert - insert rmap_item pointing to new ksm page 906 * into the stable tree. 907 * 908 * This function returns the stable tree node just allocated on success, 909 * NULL otherwise. 910 */ 911static struct stable_node *stable_tree_insert(struct page *kpage) 912{ 913 struct rb_node **new = &root_stable_tree.rb_node; 914 struct rb_node *parent = NULL; 915 struct stable_node *stable_node; 916 917 while (*new) { 918 int ret; 919 920 cond_resched(); 921 stable_node = rb_entry(*new, struct stable_node, node); 922 923 ret = memcmp_pages(kpage, stable_node->page); 924 925 parent = *new; 926 if (ret < 0) 927 new = &parent->rb_left; 928 else if (ret > 0) 929 new = &parent->rb_right; 930 else { 931 /* 932 * It is not a bug that stable_tree_search() didn't 933 * find this node: because at that time our page was 934 * not yet write-protected, so may have changed since. 935 */ 936 return NULL; 937 } 938 } 939 940 stable_node = alloc_stable_node(); 941 if (!stable_node) 942 return NULL; 943 944 rb_link_node(&stable_node->node, parent, new); 945 rb_insert_color(&stable_node->node, &root_stable_tree); 946 947 INIT_HLIST_HEAD(&stable_node->hlist); 948 949 get_page(kpage); 950 stable_node->page = kpage; 951 set_page_stable_node(kpage, stable_node); 952 953 return stable_node; 954} 955 956/* 957 * unstable_tree_search_insert - search for identical page, 958 * else insert rmap_item into the unstable tree. 959 * 960 * This function searches for a page in the unstable tree identical to the 961 * page currently being scanned; and if no identical page is found in the 962 * tree, we insert rmap_item as a new object into the unstable tree. 963 * 964 * This function returns pointer to rmap_item found to be identical 965 * to the currently scanned page, NULL otherwise. 966 * 967 * This function does both searching and inserting, because they share 968 * the same walking algorithm in an rbtree. 969 */ 970static 971struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item, 972 struct page *page, 973 struct page **tree_pagep) 974 975{ 976 struct rb_node **new = &root_unstable_tree.rb_node; 977 struct rb_node *parent = NULL; 978 979 while (*new) { 980 struct rmap_item *tree_rmap_item; 981 struct page *tree_page; 982 int ret; 983 984 cond_resched(); 985 tree_rmap_item = rb_entry(*new, struct rmap_item, node); 986 tree_page = get_mergeable_page(tree_rmap_item); 987 if (!tree_page) 988 return NULL; 989 990 /* 991 * Don't substitute a ksm page for a forked page. 992 */ 993 if (page == tree_page) { 994 put_page(tree_page); 995 return NULL; 996 } 997 998 ret = memcmp_pages(page, tree_page); 999 1000 parent = *new; 1001 if (ret < 0) { 1002 put_page(tree_page); 1003 new = &parent->rb_left; 1004 } else if (ret > 0) { 1005 put_page(tree_page); 1006 new = &parent->rb_right; 1007 } else { 1008 *tree_pagep = tree_page; 1009 return tree_rmap_item; 1010 } 1011 } 1012 1013 rmap_item->address |= UNSTABLE_FLAG; 1014 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK); 1015 rb_link_node(&rmap_item->node, parent, new); 1016 rb_insert_color(&rmap_item->node, &root_unstable_tree); 1017 1018 ksm_pages_unshared++; 1019 return NULL; 1020} 1021 1022/* 1023 * stable_tree_append - add another rmap_item to the linked list of 1024 * rmap_items hanging off a given node of the stable tree, all sharing 1025 * the same ksm page. 1026 */ 1027static void stable_tree_append(struct rmap_item *rmap_item, 1028 struct stable_node *stable_node) 1029{ 1030 rmap_item->head = stable_node; 1031 rmap_item->address |= STABLE_FLAG; 1032 hlist_add_head(&rmap_item->hlist, &stable_node->hlist); 1033 1034 if (rmap_item->hlist.next) 1035 ksm_pages_sharing++; 1036 else 1037 ksm_pages_shared++; 1038} 1039 1040/* 1041 * cmp_and_merge_page - first see if page can be merged into the stable tree; 1042 * if not, compare checksum to previous and if it's the same, see if page can 1043 * be inserted into the unstable tree, or merged with a page already there and 1044 * both transferred to the stable tree. 1045 * 1046 * @page: the page that we are searching identical page to. 1047 * @rmap_item: the reverse mapping into the virtual address of this page 1048 */ 1049static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item) 1050{ 1051 struct rmap_item *tree_rmap_item; 1052 struct page *tree_page = NULL; 1053 struct stable_node *stable_node; 1054 struct page *kpage; 1055 unsigned int checksum; 1056 int err; 1057 1058 remove_rmap_item_from_tree(rmap_item); 1059 1060 /* We first start with searching the page inside the stable tree */ 1061 stable_node = stable_tree_search(page); 1062 if (stable_node) { 1063 kpage = stable_node->page; 1064 err = try_to_merge_with_ksm_page(rmap_item, page, kpage); 1065 if (!err) { 1066 /* 1067 * The page was successfully merged: 1068 * add its rmap_item to the stable tree. 1069 */ 1070 stable_tree_append(rmap_item, stable_node); 1071 } 1072 put_page(kpage); 1073 return; 1074 } 1075 1076 /* 1077 * A ksm page might have got here by fork, but its other 1078 * references have already been removed from the stable tree. 1079 * Or it might be left over from a break_ksm which failed 1080 * when the mem_cgroup had reached its limit: try again now. 1081 */ 1082 if (PageKsm(page)) 1083 break_cow(rmap_item); 1084 1085 /* 1086 * In case the hash value of the page was changed from the last time we 1087 * have calculated it, this page to be changed frequely, therefore we 1088 * don't want to insert it to the unstable tree, and we don't want to 1089 * waste our time to search if there is something identical to it there. 1090 */ 1091 checksum = calc_checksum(page); 1092 if (rmap_item->oldchecksum != checksum) { 1093 rmap_item->oldchecksum = checksum; 1094 return; 1095 } 1096 1097 tree_rmap_item = 1098 unstable_tree_search_insert(rmap_item, page, &tree_page); 1099 if (tree_rmap_item) { 1100 kpage = try_to_merge_two_pages(rmap_item, page, 1101 tree_rmap_item, tree_page); 1102 put_page(tree_page); 1103 /* 1104 * As soon as we merge this page, we want to remove the 1105 * rmap_item of the page we have merged with from the unstable 1106 * tree, and insert it instead as new node in the stable tree. 1107 */ 1108 if (kpage) { 1109 remove_rmap_item_from_tree(tree_rmap_item); 1110 1111 stable_node = stable_tree_insert(kpage); 1112 if (stable_node) { 1113 stable_tree_append(tree_rmap_item, stable_node); 1114 stable_tree_append(rmap_item, stable_node); 1115 } 1116 put_page(kpage); 1117 1118 /* 1119 * If we fail to insert the page into the stable tree, 1120 * we will have 2 virtual addresses that are pointing 1121 * to a ksm page left outside the stable tree, 1122 * in which case we need to break_cow on both. 1123 */ 1124 if (!stable_node) { 1125 break_cow(tree_rmap_item); 1126 break_cow(rmap_item); 1127 } 1128 } 1129 } 1130} 1131 1132static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot, 1133 struct rmap_item **rmap_list, 1134 unsigned long addr) 1135{ 1136 struct rmap_item *rmap_item; 1137 1138 while (*rmap_list) { 1139 rmap_item = *rmap_list; 1140 if ((rmap_item->address & PAGE_MASK) == addr) 1141 return rmap_item; 1142 if (rmap_item->address > addr) 1143 break; 1144 *rmap_list = rmap_item->rmap_list; 1145 remove_rmap_item_from_tree(rmap_item); 1146 free_rmap_item(rmap_item); 1147 } 1148 1149 rmap_item = alloc_rmap_item(); 1150 if (rmap_item) { 1151 /* It has already been zeroed */ 1152 rmap_item->mm = mm_slot->mm; 1153 rmap_item->address = addr; 1154 rmap_item->rmap_list = *rmap_list; 1155 *rmap_list = rmap_item; 1156 } 1157 return rmap_item; 1158} 1159 1160static struct rmap_item *scan_get_next_rmap_item(struct page **page) 1161{ 1162 struct mm_struct *mm; 1163 struct mm_slot *slot; 1164 struct vm_area_struct *vma; 1165 struct rmap_item *rmap_item; 1166 1167 if (list_empty(&ksm_mm_head.mm_list)) 1168 return NULL; 1169 1170 slot = ksm_scan.mm_slot; 1171 if (slot == &ksm_mm_head) { 1172 root_unstable_tree = RB_ROOT; 1173 1174 spin_lock(&ksm_mmlist_lock); 1175 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list); 1176 ksm_scan.mm_slot = slot; 1177 spin_unlock(&ksm_mmlist_lock); 1178next_mm: 1179 ksm_scan.address = 0; 1180 ksm_scan.rmap_list = &slot->rmap_list; 1181 } 1182 1183 mm = slot->mm; 1184 down_read(&mm->mmap_sem); 1185 if (ksm_test_exit(mm)) 1186 vma = NULL; 1187 else 1188 vma = find_vma(mm, ksm_scan.address); 1189 1190 for (; vma; vma = vma->vm_next) { 1191 if (!(vma->vm_flags & VM_MERGEABLE)) 1192 continue; 1193 if (ksm_scan.address < vma->vm_start) 1194 ksm_scan.address = vma->vm_start; 1195 if (!vma->anon_vma) 1196 ksm_scan.address = vma->vm_end; 1197 1198 while (ksm_scan.address < vma->vm_end) { 1199 if (ksm_test_exit(mm)) 1200 break; 1201 *page = follow_page(vma, ksm_scan.address, FOLL_GET); 1202 if (*page && PageAnon(*page)) { 1203 flush_anon_page(vma, *page, ksm_scan.address); 1204 flush_dcache_page(*page); 1205 rmap_item = get_next_rmap_item(slot, 1206 ksm_scan.rmap_list, ksm_scan.address); 1207 if (rmap_item) { 1208 ksm_scan.rmap_list = 1209 &rmap_item->rmap_list; 1210 ksm_scan.address += PAGE_SIZE; 1211 } else 1212 put_page(*page); 1213 up_read(&mm->mmap_sem); 1214 return rmap_item; 1215 } 1216 if (*page) 1217 put_page(*page); 1218 ksm_scan.address += PAGE_SIZE; 1219 cond_resched(); 1220 } 1221 } 1222 1223 if (ksm_test_exit(mm)) { 1224 ksm_scan.address = 0; 1225 ksm_scan.rmap_list = &slot->rmap_list; 1226 } 1227 /* 1228 * Nuke all the rmap_items that are above this current rmap: 1229 * because there were no VM_MERGEABLE vmas with such addresses. 1230 */ 1231 remove_trailing_rmap_items(slot, ksm_scan.rmap_list); 1232 1233 spin_lock(&ksm_mmlist_lock); 1234 ksm_scan.mm_slot = list_entry(slot->mm_list.next, 1235 struct mm_slot, mm_list); 1236 if (ksm_scan.address == 0) { 1237 /* 1238 * We've completed a full scan of all vmas, holding mmap_sem 1239 * throughout, and found no VM_MERGEABLE: so do the same as 1240 * __ksm_exit does to remove this mm from all our lists now. 1241 * This applies either when cleaning up after __ksm_exit 1242 * (but beware: we can reach here even before __ksm_exit), 1243 * or when all VM_MERGEABLE areas have been unmapped (and 1244 * mmap_sem then protects against race with MADV_MERGEABLE). 1245 */ 1246 hlist_del(&slot->link); 1247 list_del(&slot->mm_list); 1248 spin_unlock(&ksm_mmlist_lock); 1249 1250 free_mm_slot(slot); 1251 clear_bit(MMF_VM_MERGEABLE, &mm->flags); 1252 up_read(&mm->mmap_sem); 1253 mmdrop(mm); 1254 } else { 1255 spin_unlock(&ksm_mmlist_lock); 1256 up_read(&mm->mmap_sem); 1257 } 1258 1259 /* Repeat until we've completed scanning the whole list */ 1260 slot = ksm_scan.mm_slot; 1261 if (slot != &ksm_mm_head) 1262 goto next_mm; 1263 1264 ksm_scan.seqnr++; 1265 return NULL; 1266} 1267 1268/** 1269 * ksm_do_scan - the ksm scanner main worker function. 1270 * @scan_npages - number of pages we want to scan before we return. 1271 */ 1272static void ksm_do_scan(unsigned int scan_npages) 1273{ 1274 struct rmap_item *rmap_item; 1275 struct page *page; 1276 1277 while (scan_npages--) { 1278 cond_resched(); 1279 rmap_item = scan_get_next_rmap_item(&page); 1280 if (!rmap_item) 1281 return; 1282 if (!PageKsm(page) || !in_stable_tree(rmap_item)) 1283 cmp_and_merge_page(page, rmap_item); 1284 else if (page_mapcount(page) == 1) { 1285 /* 1286 * Replace now-unshared ksm page by ordinary page. 1287 */ 1288 break_cow(rmap_item); 1289 remove_rmap_item_from_tree(rmap_item); 1290 rmap_item->oldchecksum = calc_checksum(page); 1291 } 1292 put_page(page); 1293 } 1294} 1295 1296static int ksmd_should_run(void) 1297{ 1298 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list); 1299} 1300 1301static int ksm_scan_thread(void *nothing) 1302{ 1303 set_user_nice(current, 5); 1304 1305 while (!kthread_should_stop()) { 1306 mutex_lock(&ksm_thread_mutex); 1307 if (ksmd_should_run()) 1308 ksm_do_scan(ksm_thread_pages_to_scan); 1309 mutex_unlock(&ksm_thread_mutex); 1310 1311 if (ksmd_should_run()) { 1312 schedule_timeout_interruptible( 1313 msecs_to_jiffies(ksm_thread_sleep_millisecs)); 1314 } else { 1315 wait_event_interruptible(ksm_thread_wait, 1316 ksmd_should_run() || kthread_should_stop()); 1317 } 1318 } 1319 return 0; 1320} 1321 1322int ksm_madvise(struct vm_area_struct *vma, unsigned long start, 1323 unsigned long end, int advice, unsigned long *vm_flags) 1324{ 1325 struct mm_struct *mm = vma->vm_mm; 1326 int err; 1327 1328 switch (advice) { 1329 case MADV_MERGEABLE: 1330 /* 1331 * Be somewhat over-protective for now! 1332 */ 1333 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE | 1334 VM_PFNMAP | VM_IO | VM_DONTEXPAND | 1335 VM_RESERVED | VM_HUGETLB | VM_INSERTPAGE | 1336 VM_MIXEDMAP | VM_SAO)) 1337 return 0; /* just ignore the advice */ 1338 1339 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) { 1340 err = __ksm_enter(mm); 1341 if (err) 1342 return err; 1343 } 1344 1345 *vm_flags |= VM_MERGEABLE; 1346 break; 1347 1348 case MADV_UNMERGEABLE: 1349 if (!(*vm_flags & VM_MERGEABLE)) 1350 return 0; /* just ignore the advice */ 1351 1352 if (vma->anon_vma) { 1353 err = unmerge_ksm_pages(vma, start, end); 1354 if (err) 1355 return err; 1356 } 1357 1358 *vm_flags &= ~VM_MERGEABLE; 1359 break; 1360 } 1361 1362 return 0; 1363} 1364 1365int __ksm_enter(struct mm_struct *mm) 1366{ 1367 struct mm_slot *mm_slot; 1368 int needs_wakeup; 1369 1370 mm_slot = alloc_mm_slot(); 1371 if (!mm_slot) 1372 return -ENOMEM; 1373 1374 /* Check ksm_run too? Would need tighter locking */ 1375 needs_wakeup = list_empty(&ksm_mm_head.mm_list); 1376 1377 spin_lock(&ksm_mmlist_lock); 1378 insert_to_mm_slots_hash(mm, mm_slot); 1379 /* 1380 * Insert just behind the scanning cursor, to let the area settle 1381 * down a little; when fork is followed by immediate exec, we don't 1382 * want ksmd to waste time setting up and tearing down an rmap_list. 1383 */ 1384 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list); 1385 spin_unlock(&ksm_mmlist_lock); 1386 1387 set_bit(MMF_VM_MERGEABLE, &mm->flags); 1388 atomic_inc(&mm->mm_count); 1389 1390 if (needs_wakeup) 1391 wake_up_interruptible(&ksm_thread_wait); 1392 1393 return 0; 1394} 1395 1396void __ksm_exit(struct mm_struct *mm) 1397{ 1398 struct mm_slot *mm_slot; 1399 int easy_to_free = 0; 1400 1401 /* 1402 * This process is exiting: if it's straightforward (as is the 1403 * case when ksmd was never running), free mm_slot immediately. 1404 * But if it's at the cursor or has rmap_items linked to it, use 1405 * mmap_sem to synchronize with any break_cows before pagetables 1406 * are freed, and leave the mm_slot on the list for ksmd to free. 1407 * Beware: ksm may already have noticed it exiting and freed the slot. 1408 */ 1409 1410 spin_lock(&ksm_mmlist_lock); 1411 mm_slot = get_mm_slot(mm); 1412 if (mm_slot && ksm_scan.mm_slot != mm_slot) { 1413 if (!mm_slot->rmap_list) { 1414 hlist_del(&mm_slot->link); 1415 list_del(&mm_slot->mm_list); 1416 easy_to_free = 1; 1417 } else { 1418 list_move(&mm_slot->mm_list, 1419 &ksm_scan.mm_slot->mm_list); 1420 } 1421 } 1422 spin_unlock(&ksm_mmlist_lock); 1423 1424 if (easy_to_free) { 1425 free_mm_slot(mm_slot); 1426 clear_bit(MMF_VM_MERGEABLE, &mm->flags); 1427 mmdrop(mm); 1428 } else if (mm_slot) { 1429 down_write(&mm->mmap_sem); 1430 up_write(&mm->mmap_sem); 1431 } 1432} 1433 1434#ifdef CONFIG_SYSFS 1435/* 1436 * This all compiles without CONFIG_SYSFS, but is a waste of space. 1437 */ 1438 1439#define KSM_ATTR_RO(_name) \ 1440 static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 1441#define KSM_ATTR(_name) \ 1442 static struct kobj_attribute _name##_attr = \ 1443 __ATTR(_name, 0644, _name##_show, _name##_store) 1444 1445static ssize_t sleep_millisecs_show(struct kobject *kobj, 1446 struct kobj_attribute *attr, char *buf) 1447{ 1448 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs); 1449} 1450 1451static ssize_t sleep_millisecs_store(struct kobject *kobj, 1452 struct kobj_attribute *attr, 1453 const char *buf, size_t count) 1454{ 1455 unsigned long msecs; 1456 int err; 1457 1458 err = strict_strtoul(buf, 10, &msecs); 1459 if (err || msecs > UINT_MAX) 1460 return -EINVAL; 1461 1462 ksm_thread_sleep_millisecs = msecs; 1463 1464 return count; 1465} 1466KSM_ATTR(sleep_millisecs); 1467 1468static ssize_t pages_to_scan_show(struct kobject *kobj, 1469 struct kobj_attribute *attr, char *buf) 1470{ 1471 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan); 1472} 1473 1474static ssize_t pages_to_scan_store(struct kobject *kobj, 1475 struct kobj_attribute *attr, 1476 const char *buf, size_t count) 1477{ 1478 int err; 1479 unsigned long nr_pages; 1480 1481 err = strict_strtoul(buf, 10, &nr_pages); 1482 if (err || nr_pages > UINT_MAX) 1483 return -EINVAL; 1484 1485 ksm_thread_pages_to_scan = nr_pages; 1486 1487 return count; 1488} 1489KSM_ATTR(pages_to_scan); 1490 1491static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr, 1492 char *buf) 1493{ 1494 return sprintf(buf, "%u\n", ksm_run); 1495} 1496 1497static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr, 1498 const char *buf, size_t count) 1499{ 1500 int err; 1501 unsigned long flags; 1502 1503 err = strict_strtoul(buf, 10, &flags); 1504 if (err || flags > UINT_MAX) 1505 return -EINVAL; 1506 if (flags > KSM_RUN_UNMERGE) 1507 return -EINVAL; 1508 1509 /* 1510 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running. 1511 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items, 1512 * breaking COW to free the unswappable pages_shared (but leaves 1513 * mm_slots on the list for when ksmd may be set running again). 1514 */ 1515 1516 mutex_lock(&ksm_thread_mutex); 1517 if (ksm_run != flags) { 1518 ksm_run = flags; 1519 if (flags & KSM_RUN_UNMERGE) { 1520 current->flags |= PF_OOM_ORIGIN; 1521 err = unmerge_and_remove_all_rmap_items(); 1522 current->flags &= ~PF_OOM_ORIGIN; 1523 if (err) { 1524 ksm_run = KSM_RUN_STOP; 1525 count = err; 1526 } 1527 } 1528 } 1529 mutex_unlock(&ksm_thread_mutex); 1530 1531 if (flags & KSM_RUN_MERGE) 1532 wake_up_interruptible(&ksm_thread_wait); 1533 1534 return count; 1535} 1536KSM_ATTR(run); 1537 1538static ssize_t max_kernel_pages_store(struct kobject *kobj, 1539 struct kobj_attribute *attr, 1540 const char *buf, size_t count) 1541{ 1542 int err; 1543 unsigned long nr_pages; 1544 1545 err = strict_strtoul(buf, 10, &nr_pages); 1546 if (err) 1547 return -EINVAL; 1548 1549 ksm_max_kernel_pages = nr_pages; 1550 1551 return count; 1552} 1553 1554static ssize_t max_kernel_pages_show(struct kobject *kobj, 1555 struct kobj_attribute *attr, char *buf) 1556{ 1557 return sprintf(buf, "%lu\n", ksm_max_kernel_pages); 1558} 1559KSM_ATTR(max_kernel_pages); 1560 1561static ssize_t pages_shared_show(struct kobject *kobj, 1562 struct kobj_attribute *attr, char *buf) 1563{ 1564 return sprintf(buf, "%lu\n", ksm_pages_shared); 1565} 1566KSM_ATTR_RO(pages_shared); 1567 1568static ssize_t pages_sharing_show(struct kobject *kobj, 1569 struct kobj_attribute *attr, char *buf) 1570{ 1571 return sprintf(buf, "%lu\n", ksm_pages_sharing); 1572} 1573KSM_ATTR_RO(pages_sharing); 1574 1575static ssize_t pages_unshared_show(struct kobject *kobj, 1576 struct kobj_attribute *attr, char *buf) 1577{ 1578 return sprintf(buf, "%lu\n", ksm_pages_unshared); 1579} 1580KSM_ATTR_RO(pages_unshared); 1581 1582static ssize_t pages_volatile_show(struct kobject *kobj, 1583 struct kobj_attribute *attr, char *buf) 1584{ 1585 long ksm_pages_volatile; 1586 1587 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared 1588 - ksm_pages_sharing - ksm_pages_unshared; 1589 /* 1590 * It was not worth any locking to calculate that statistic, 1591 * but it might therefore sometimes be negative: conceal that. 1592 */ 1593 if (ksm_pages_volatile < 0) 1594 ksm_pages_volatile = 0; 1595 return sprintf(buf, "%ld\n", ksm_pages_volatile); 1596} 1597KSM_ATTR_RO(pages_volatile); 1598 1599static ssize_t full_scans_show(struct kobject *kobj, 1600 struct kobj_attribute *attr, char *buf) 1601{ 1602 return sprintf(buf, "%lu\n", ksm_scan.seqnr); 1603} 1604KSM_ATTR_RO(full_scans); 1605 1606static struct attribute *ksm_attrs[] = { 1607 &sleep_millisecs_attr.attr, 1608 &pages_to_scan_attr.attr, 1609 &run_attr.attr, 1610 &max_kernel_pages_attr.attr, 1611 &pages_shared_attr.attr, 1612 &pages_sharing_attr.attr, 1613 &pages_unshared_attr.attr, 1614 &pages_volatile_attr.attr, 1615 &full_scans_attr.attr, 1616 NULL, 1617}; 1618 1619static struct attribute_group ksm_attr_group = { 1620 .attrs = ksm_attrs, 1621 .name = "ksm", 1622}; 1623#endif /* CONFIG_SYSFS */ 1624 1625static int __init ksm_init(void) 1626{ 1627 struct task_struct *ksm_thread; 1628 int err; 1629 1630 ksm_max_kernel_pages = totalram_pages / 4; 1631 1632 err = ksm_slab_init(); 1633 if (err) 1634 goto out; 1635 1636 err = mm_slots_hash_init(); 1637 if (err) 1638 goto out_free1; 1639 1640 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd"); 1641 if (IS_ERR(ksm_thread)) { 1642 printk(KERN_ERR "ksm: creating kthread failed\n"); 1643 err = PTR_ERR(ksm_thread); 1644 goto out_free2; 1645 } 1646 1647#ifdef CONFIG_SYSFS 1648 err = sysfs_create_group(mm_kobj, &ksm_attr_group); 1649 if (err) { 1650 printk(KERN_ERR "ksm: register sysfs failed\n"); 1651 kthread_stop(ksm_thread); 1652 goto out_free2; 1653 } 1654#else 1655 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */ 1656 1657#endif /* CONFIG_SYSFS */ 1658 1659 return 0; 1660 1661out_free2: 1662 mm_slots_hash_free(); 1663out_free1: 1664 ksm_slab_free(); 1665out: 1666 return err; 1667} 1668module_init(ksm_init) 1669