| 1 | // SPDX-License-Identifier: GPL-2.0-only |
| 2 | /* |
| 3 | * Memory merging support. |
| 4 | * |
| 5 | * This code enables dynamic sharing of identical pages found in different |
| 6 | * memory areas, even if they are not shared by fork() |
| 7 | * |
| 8 | * Copyright (C) 2008-2009 Red Hat, Inc. |
| 9 | * Authors: |
| 10 | * Izik Eidus |
| 11 | * Andrea Arcangeli |
| 12 | * Chris Wright |
| 13 | * Hugh Dickins |
| 14 | */ |
| 15 | |
| 16 | #include <linux/errno.h> |
| 17 | #include <linux/mm.h> |
| 18 | #include <linux/mm_inline.h> |
| 19 | #include <linux/fs.h> |
| 20 | #include <linux/mman.h> |
| 21 | #include <linux/sched.h> |
| 22 | #include <linux/sched/mm.h> |
| 23 | #include <linux/sched/coredump.h> |
| 24 | #include <linux/rwsem.h> |
| 25 | #include <linux/pagemap.h> |
| 26 | #include <linux/rmap.h> |
| 27 | #include <linux/spinlock.h> |
| 28 | #include <linux/xxhash.h> |
| 29 | #include <linux/delay.h> |
| 30 | #include <linux/kthread.h> |
| 31 | #include <linux/wait.h> |
| 32 | #include <linux/slab.h> |
| 33 | #include <linux/rbtree.h> |
| 34 | #include <linux/memory.h> |
| 35 | #include <linux/mmu_notifier.h> |
| 36 | #include <linux/swap.h> |
| 37 | #include <linux/ksm.h> |
| 38 | #include <linux/hashtable.h> |
| 39 | #include <linux/freezer.h> |
| 40 | #include <linux/oom.h> |
| 41 | #include <linux/numa.h> |
| 42 | #include <linux/pagewalk.h> |
| 43 | |
| 44 | #include <asm/tlbflush.h> |
| 45 | #include "internal.h" |
| 46 | #include "mm_slot.h" |
| 47 | |
| 48 | #define CREATE_TRACE_POINTS |
| 49 | #include <trace/events/ksm.h> |
| 50 | |
| 51 | #ifdef CONFIG_NUMA |
| 52 | #define NUMA(x) (x) |
| 53 | #define DO_NUMA(x) do { (x); } while (0) |
| 54 | #else |
| 55 | #define NUMA(x) (0) |
| 56 | #define DO_NUMA(x) do { } while (0) |
| 57 | #endif |
| 58 | |
| 59 | /** |
| 60 | * DOC: Overview |
| 61 | * |
| 62 | * A few notes about the KSM scanning process, |
| 63 | * to make it easier to understand the data structures below: |
| 64 | * |
| 65 | * In order to reduce excessive scanning, KSM sorts the memory pages by their |
| 66 | * contents into a data structure that holds pointers to the pages' locations. |
| 67 | * |
| 68 | * Since the contents of the pages may change at any moment, KSM cannot just |
| 69 | * insert the pages into a normal sorted tree and expect it to find anything. |
| 70 | * Therefore KSM uses two data structures - the stable and the unstable tree. |
| 71 | * |
| 72 | * The stable tree holds pointers to all the merged pages (ksm pages), sorted |
| 73 | * by their contents. Because each such page is write-protected, searching on |
| 74 | * this tree is fully assured to be working (except when pages are unmapped), |
| 75 | * and therefore this tree is called the stable tree. |
| 76 | * |
| 77 | * The stable tree node includes information required for reverse |
| 78 | * mapping from a KSM page to virtual addresses that map this page. |
| 79 | * |
| 80 | * In order to avoid large latencies of the rmap walks on KSM pages, |
| 81 | * KSM maintains two types of nodes in the stable tree: |
| 82 | * |
| 83 | * * the regular nodes that keep the reverse mapping structures in a |
| 84 | * linked list |
| 85 | * * the "chains" that link nodes ("dups") that represent the same |
| 86 | * write protected memory content, but each "dup" corresponds to a |
| 87 | * different KSM page copy of that content |
| 88 | * |
| 89 | * Internally, the regular nodes, "dups" and "chains" are represented |
| 90 | * using the same struct ksm_stable_node structure. |
| 91 | * |
| 92 | * In addition to the stable tree, KSM uses a second data structure called the |
| 93 | * unstable tree: this tree holds pointers to pages which have been found to |
| 94 | * be "unchanged for a period of time". The unstable tree sorts these pages |
| 95 | * by their contents, but since they are not write-protected, KSM cannot rely |
| 96 | * upon the unstable tree to work correctly - the unstable tree is liable to |
| 97 | * be corrupted as its contents are modified, and so it is called unstable. |
| 98 | * |
| 99 | * KSM solves this problem by several techniques: |
| 100 | * |
| 101 | * 1) The unstable tree is flushed every time KSM completes scanning all |
| 102 | * memory areas, and then the tree is rebuilt again from the beginning. |
| 103 | * 2) KSM will only insert into the unstable tree, pages whose hash value |
| 104 | * has not changed since the previous scan of all memory areas. |
| 105 | * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the |
| 106 | * colors of the nodes and not on their contents, assuring that even when |
| 107 | * the tree gets "corrupted" it won't get out of balance, so scanning time |
| 108 | * remains the same (also, searching and inserting nodes in an rbtree uses |
| 109 | * the same algorithm, so we have no overhead when we flush and rebuild). |
| 110 | * 4) KSM never flushes the stable tree, which means that even if it were to |
| 111 | * take 10 attempts to find a page in the unstable tree, once it is found, |
| 112 | * it is secured in the stable tree. (When we scan a new page, we first |
| 113 | * compare it against the stable tree, and then against the unstable tree.) |
| 114 | * |
| 115 | * If the merge_across_nodes tunable is unset, then KSM maintains multiple |
| 116 | * stable trees and multiple unstable trees: one of each for each NUMA node. |
| 117 | */ |
| 118 | |
| 119 | /** |
| 120 | * struct ksm_mm_slot - ksm information per mm that is being scanned |
| 121 | * @slot: hash lookup from mm to mm_slot |
| 122 | * @rmap_list: head for this mm_slot's singly-linked list of rmap_items |
| 123 | */ |
| 124 | struct ksm_mm_slot { |
| 125 | struct mm_slot slot; |
| 126 | struct ksm_rmap_item *rmap_list; |
| 127 | }; |
| 128 | |
| 129 | /** |
| 130 | * struct ksm_scan - cursor for scanning |
| 131 | * @mm_slot: the current mm_slot we are scanning |
| 132 | * @address: the next address inside that to be scanned |
| 133 | * @rmap_list: link to the next rmap to be scanned in the rmap_list |
| 134 | * @seqnr: count of completed full scans (needed when removing unstable node) |
| 135 | * |
| 136 | * There is only the one ksm_scan instance of this cursor structure. |
| 137 | */ |
| 138 | struct ksm_scan { |
| 139 | struct ksm_mm_slot *mm_slot; |
| 140 | unsigned long address; |
| 141 | struct ksm_rmap_item **rmap_list; |
| 142 | unsigned long seqnr; |
| 143 | }; |
| 144 | |
| 145 | /** |
| 146 | * struct ksm_stable_node - node of the stable rbtree |
| 147 | * @node: rb node of this ksm page in the stable tree |
| 148 | * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list |
| 149 | * @hlist_dup: linked into the stable_node->hlist with a stable_node chain |
| 150 | * @list: linked into migrate_nodes, pending placement in the proper node tree |
| 151 | * @hlist: hlist head of rmap_items using this ksm page |
| 152 | * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid) |
| 153 | * @chain_prune_time: time of the last full garbage collection |
| 154 | * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN |
| 155 | * @nid: NUMA node id of stable tree in which linked (may not match kpfn) |
| 156 | */ |
| 157 | struct ksm_stable_node { |
| 158 | union { |
| 159 | struct rb_node node; /* when node of stable tree */ |
| 160 | struct { /* when listed for migration */ |
| 161 | struct list_head *head; |
| 162 | struct { |
| 163 | struct hlist_node hlist_dup; |
| 164 | struct list_head list; |
| 165 | }; |
| 166 | }; |
| 167 | }; |
| 168 | struct hlist_head hlist; |
| 169 | union { |
| 170 | unsigned long kpfn; |
| 171 | unsigned long chain_prune_time; |
| 172 | }; |
| 173 | /* |
| 174 | * STABLE_NODE_CHAIN can be any negative number in |
| 175 | * rmap_hlist_len negative range, but better not -1 to be able |
| 176 | * to reliably detect underflows. |
| 177 | */ |
| 178 | #define STABLE_NODE_CHAIN -1024 |
| 179 | int rmap_hlist_len; |
| 180 | #ifdef CONFIG_NUMA |
| 181 | int nid; |
| 182 | #endif |
| 183 | }; |
| 184 | |
| 185 | /** |
| 186 | * struct ksm_rmap_item - reverse mapping item for virtual addresses |
| 187 | * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list |
| 188 | * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree |
| 189 | * @nid: NUMA node id of unstable tree in which linked (may not match page) |
| 190 | * @mm: the memory structure this rmap_item is pointing into |
| 191 | * @address: the virtual address this rmap_item tracks (+ flags in low bits) |
| 192 | * @oldchecksum: previous checksum of the page at that virtual address |
| 193 | * @node: rb node of this rmap_item in the unstable tree |
| 194 | * @head: pointer to stable_node heading this list in the stable tree |
| 195 | * @hlist: link into hlist of rmap_items hanging off that stable_node |
| 196 | */ |
| 197 | struct ksm_rmap_item { |
| 198 | struct ksm_rmap_item *rmap_list; |
| 199 | union { |
| 200 | struct anon_vma *anon_vma; /* when stable */ |
| 201 | #ifdef CONFIG_NUMA |
| 202 | int nid; /* when node of unstable tree */ |
| 203 | #endif |
| 204 | }; |
| 205 | struct mm_struct *mm; |
| 206 | unsigned long address; /* + low bits used for flags below */ |
| 207 | unsigned int oldchecksum; /* when unstable */ |
| 208 | union { |
| 209 | struct rb_node node; /* when node of unstable tree */ |
| 210 | struct { /* when listed from stable tree */ |
| 211 | struct ksm_stable_node *head; |
| 212 | struct hlist_node hlist; |
| 213 | }; |
| 214 | }; |
| 215 | }; |
| 216 | |
| 217 | #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */ |
| 218 | #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */ |
| 219 | #define STABLE_FLAG 0x200 /* is listed from the stable tree */ |
| 220 | |
| 221 | /* The stable and unstable tree heads */ |
| 222 | static struct rb_root one_stable_tree[1] = { RB_ROOT }; |
| 223 | static struct rb_root one_unstable_tree[1] = { RB_ROOT }; |
| 224 | static struct rb_root *root_stable_tree = one_stable_tree; |
| 225 | static struct rb_root *root_unstable_tree = one_unstable_tree; |
| 226 | |
| 227 | /* Recently migrated nodes of stable tree, pending proper placement */ |
| 228 | static LIST_HEAD(migrate_nodes); |
| 229 | #define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev) |
| 230 | |
| 231 | #define MM_SLOTS_HASH_BITS 10 |
| 232 | static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); |
| 233 | |
| 234 | static struct ksm_mm_slot ksm_mm_head = { |
| 235 | .slot.mm_node = LIST_HEAD_INIT(ksm_mm_head.slot.mm_node), |
| 236 | }; |
| 237 | static struct ksm_scan ksm_scan = { |
| 238 | .mm_slot = &ksm_mm_head, |
| 239 | }; |
| 240 | |
| 241 | static struct kmem_cache *rmap_item_cache; |
| 242 | static struct kmem_cache *stable_node_cache; |
| 243 | static struct kmem_cache *mm_slot_cache; |
| 244 | |
| 245 | /* The number of nodes in the stable tree */ |
| 246 | static unsigned long ksm_pages_shared; |
| 247 | |
| 248 | /* The number of page slots additionally sharing those nodes */ |
| 249 | static unsigned long ksm_pages_sharing; |
| 250 | |
| 251 | /* The number of nodes in the unstable tree */ |
| 252 | static unsigned long ksm_pages_unshared; |
| 253 | |
| 254 | /* The number of rmap_items in use: to calculate pages_volatile */ |
| 255 | static unsigned long ksm_rmap_items; |
| 256 | |
| 257 | /* The number of stable_node chains */ |
| 258 | static unsigned long ksm_stable_node_chains; |
| 259 | |
| 260 | /* The number of stable_node dups linked to the stable_node chains */ |
| 261 | static unsigned long ksm_stable_node_dups; |
| 262 | |
| 263 | /* Delay in pruning stale stable_node_dups in the stable_node_chains */ |
| 264 | static unsigned int ksm_stable_node_chains_prune_millisecs = 2000; |
| 265 | |
| 266 | /* Maximum number of page slots sharing a stable node */ |
| 267 | static int ksm_max_page_sharing = 256; |
| 268 | |
| 269 | /* Number of pages ksmd should scan in one batch */ |
| 270 | static unsigned int ksm_thread_pages_to_scan = 100; |
| 271 | |
| 272 | /* Milliseconds ksmd should sleep between batches */ |
| 273 | static unsigned int ksm_thread_sleep_millisecs = 20; |
| 274 | |
| 275 | /* Checksum of an empty (zeroed) page */ |
| 276 | static unsigned int zero_checksum __read_mostly; |
| 277 | |
| 278 | /* Whether to merge empty (zeroed) pages with actual zero pages */ |
| 279 | static bool ksm_use_zero_pages __read_mostly; |
| 280 | |
| 281 | #ifdef CONFIG_NUMA |
| 282 | /* Zeroed when merging across nodes is not allowed */ |
| 283 | static unsigned int ksm_merge_across_nodes = 1; |
| 284 | static int ksm_nr_node_ids = 1; |
| 285 | #else |
| 286 | #define ksm_merge_across_nodes 1U |
| 287 | #define ksm_nr_node_ids 1 |
| 288 | #endif |
| 289 | |
| 290 | #define KSM_RUN_STOP 0 |
| 291 | #define KSM_RUN_MERGE 1 |
| 292 | #define KSM_RUN_UNMERGE 2 |
| 293 | #define KSM_RUN_OFFLINE 4 |
| 294 | static unsigned long ksm_run = KSM_RUN_STOP; |
| 295 | static void wait_while_offlining(void); |
| 296 | |
| 297 | static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait); |
| 298 | static DECLARE_WAIT_QUEUE_HEAD(ksm_iter_wait); |
| 299 | static DEFINE_MUTEX(ksm_thread_mutex); |
| 300 | static DEFINE_SPINLOCK(ksm_mmlist_lock); |
| 301 | |
| 302 | #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create(#__struct,\ |
| 303 | sizeof(struct __struct), __alignof__(struct __struct),\ |
| 304 | (__flags), NULL) |
| 305 | |
| 306 | static int __init ksm_slab_init(void) |
| 307 | { |
| 308 | rmap_item_cache = KSM_KMEM_CACHE(ksm_rmap_item, 0); |
| 309 | if (!rmap_item_cache) |
| 310 | goto out; |
| 311 | |
| 312 | stable_node_cache = KSM_KMEM_CACHE(ksm_stable_node, 0); |
| 313 | if (!stable_node_cache) |
| 314 | goto out_free1; |
| 315 | |
| 316 | mm_slot_cache = KSM_KMEM_CACHE(ksm_mm_slot, 0); |
| 317 | if (!mm_slot_cache) |
| 318 | goto out_free2; |
| 319 | |
| 320 | return 0; |
| 321 | |
| 322 | out_free2: |
| 323 | kmem_cache_destroy(stable_node_cache); |
| 324 | out_free1: |
| 325 | kmem_cache_destroy(rmap_item_cache); |
| 326 | out: |
| 327 | return -ENOMEM; |
| 328 | } |
| 329 | |
| 330 | static void __init ksm_slab_free(void) |
| 331 | { |
| 332 | kmem_cache_destroy(mm_slot_cache); |
| 333 | kmem_cache_destroy(stable_node_cache); |
| 334 | kmem_cache_destroy(rmap_item_cache); |
| 335 | mm_slot_cache = NULL; |
| 336 | } |
| 337 | |
| 338 | static __always_inline bool is_stable_node_chain(struct ksm_stable_node *chain) |
| 339 | { |
| 340 | return chain->rmap_hlist_len == STABLE_NODE_CHAIN; |
| 341 | } |
| 342 | |
| 343 | static __always_inline bool is_stable_node_dup(struct ksm_stable_node *dup) |
| 344 | { |
| 345 | return dup->head == STABLE_NODE_DUP_HEAD; |
| 346 | } |
| 347 | |
| 348 | static inline void stable_node_chain_add_dup(struct ksm_stable_node *dup, |
| 349 | struct ksm_stable_node *chain) |
| 350 | { |
| 351 | VM_BUG_ON(is_stable_node_dup(dup)); |
| 352 | dup->head = STABLE_NODE_DUP_HEAD; |
| 353 | VM_BUG_ON(!is_stable_node_chain(chain)); |
| 354 | hlist_add_head(&dup->hlist_dup, &chain->hlist); |
| 355 | ksm_stable_node_dups++; |
| 356 | } |
| 357 | |
| 358 | static inline void __stable_node_dup_del(struct ksm_stable_node *dup) |
| 359 | { |
| 360 | VM_BUG_ON(!is_stable_node_dup(dup)); |
| 361 | hlist_del(&dup->hlist_dup); |
| 362 | ksm_stable_node_dups--; |
| 363 | } |
| 364 | |
| 365 | static inline void stable_node_dup_del(struct ksm_stable_node *dup) |
| 366 | { |
| 367 | VM_BUG_ON(is_stable_node_chain(dup)); |
| 368 | if (is_stable_node_dup(dup)) |
| 369 | __stable_node_dup_del(dup); |
| 370 | else |
| 371 | rb_erase(&dup->node, root_stable_tree + NUMA(dup->nid)); |
| 372 | #ifdef CONFIG_DEBUG_VM |
| 373 | dup->head = NULL; |
| 374 | #endif |
| 375 | } |
| 376 | |
| 377 | static inline struct ksm_rmap_item *alloc_rmap_item(void) |
| 378 | { |
| 379 | struct ksm_rmap_item *rmap_item; |
| 380 | |
| 381 | rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL | |
| 382 | __GFP_NORETRY | __GFP_NOWARN); |
| 383 | if (rmap_item) |
| 384 | ksm_rmap_items++; |
| 385 | return rmap_item; |
| 386 | } |
| 387 | |
| 388 | static inline void free_rmap_item(struct ksm_rmap_item *rmap_item) |
| 389 | { |
| 390 | ksm_rmap_items--; |
| 391 | rmap_item->mm->ksm_rmap_items--; |
| 392 | rmap_item->mm = NULL; /* debug safety */ |
| 393 | kmem_cache_free(rmap_item_cache, rmap_item); |
| 394 | } |
| 395 | |
| 396 | static inline struct ksm_stable_node *alloc_stable_node(void) |
| 397 | { |
| 398 | /* |
| 399 | * The allocation can take too long with GFP_KERNEL when memory is under |
| 400 | * pressure, which may lead to hung task warnings. Adding __GFP_HIGH |
| 401 | * grants access to memory reserves, helping to avoid this problem. |
| 402 | */ |
| 403 | return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH); |
| 404 | } |
| 405 | |
| 406 | static inline void free_stable_node(struct ksm_stable_node *stable_node) |
| 407 | { |
| 408 | VM_BUG_ON(stable_node->rmap_hlist_len && |
| 409 | !is_stable_node_chain(stable_node)); |
| 410 | kmem_cache_free(stable_node_cache, stable_node); |
| 411 | } |
| 412 | |
| 413 | /* |
| 414 | * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's |
| 415 | * page tables after it has passed through ksm_exit() - which, if necessary, |
| 416 | * takes mmap_lock briefly to serialize against them. ksm_exit() does not set |
| 417 | * a special flag: they can just back out as soon as mm_users goes to zero. |
| 418 | * ksm_test_exit() is used throughout to make this test for exit: in some |
| 419 | * places for correctness, in some places just to avoid unnecessary work. |
| 420 | */ |
| 421 | static inline bool ksm_test_exit(struct mm_struct *mm) |
| 422 | { |
| 423 | return atomic_read(&mm->mm_users) == 0; |
| 424 | } |
| 425 | |
| 426 | static int break_ksm_pmd_entry(pmd_t *pmd, unsigned long addr, unsigned long next, |
| 427 | struct mm_walk *walk) |
| 428 | { |
| 429 | struct page *page = NULL; |
| 430 | spinlock_t *ptl; |
| 431 | pte_t *pte; |
| 432 | int ret; |
| 433 | |
| 434 | if (pmd_leaf(*pmd) || !pmd_present(*pmd)) |
| 435 | return 0; |
| 436 | |
| 437 | pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl); |
| 438 | if (pte_present(*pte)) { |
| 439 | page = vm_normal_page(walk->vma, addr, *pte); |
| 440 | } else if (!pte_none(*pte)) { |
| 441 | swp_entry_t entry = pte_to_swp_entry(*pte); |
| 442 | |
| 443 | /* |
| 444 | * As KSM pages remain KSM pages until freed, no need to wait |
| 445 | * here for migration to end. |
| 446 | */ |
| 447 | if (is_migration_entry(entry)) |
| 448 | page = pfn_swap_entry_to_page(entry); |
| 449 | } |
| 450 | ret = page && PageKsm(page); |
| 451 | pte_unmap_unlock(pte, ptl); |
| 452 | return ret; |
| 453 | } |
| 454 | |
| 455 | static const struct mm_walk_ops break_ksm_ops = { |
| 456 | .pmd_entry = break_ksm_pmd_entry, |
| 457 | }; |
| 458 | |
| 459 | /* |
| 460 | * We use break_ksm to break COW on a ksm page by triggering unsharing, |
| 461 | * such that the ksm page will get replaced by an exclusive anonymous page. |
| 462 | * |
| 463 | * We take great care only to touch a ksm page, in a VM_MERGEABLE vma, |
| 464 | * in case the application has unmapped and remapped mm,addr meanwhile. |
| 465 | * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP |
| 466 | * mmap of /dev/mem, where we would not want to touch it. |
| 467 | * |
| 468 | * FAULT_FLAG_REMOTE/FOLL_REMOTE are because we do this outside the context |
| 469 | * of the process that owns 'vma'. We also do not want to enforce |
| 470 | * protection keys here anyway. |
| 471 | */ |
| 472 | static int break_ksm(struct vm_area_struct *vma, unsigned long addr) |
| 473 | { |
| 474 | vm_fault_t ret = 0; |
| 475 | |
| 476 | do { |
| 477 | int ksm_page; |
| 478 | |
| 479 | cond_resched(); |
| 480 | ksm_page = walk_page_range_vma(vma, addr, addr + 1, |
| 481 | &break_ksm_ops, NULL); |
| 482 | if (WARN_ON_ONCE(ksm_page < 0)) |
| 483 | return ksm_page; |
| 484 | if (!ksm_page) |
| 485 | return 0; |
| 486 | ret = handle_mm_fault(vma, addr, |
| 487 | FAULT_FLAG_UNSHARE | FAULT_FLAG_REMOTE, |
| 488 | NULL); |
| 489 | } while (!(ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM))); |
| 490 | /* |
| 491 | * We must loop until we no longer find a KSM page because |
| 492 | * handle_mm_fault() may back out if there's any difficulty e.g. if |
| 493 | * pte accessed bit gets updated concurrently. |
| 494 | * |
| 495 | * VM_FAULT_SIGBUS could occur if we race with truncation of the |
| 496 | * backing file, which also invalidates anonymous pages: that's |
| 497 | * okay, that truncation will have unmapped the PageKsm for us. |
| 498 | * |
| 499 | * VM_FAULT_OOM: at the time of writing (late July 2009), setting |
| 500 | * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the |
| 501 | * current task has TIF_MEMDIE set, and will be OOM killed on return |
| 502 | * to user; and ksmd, having no mm, would never be chosen for that. |
| 503 | * |
| 504 | * But if the mm is in a limited mem_cgroup, then the fault may fail |
| 505 | * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and |
| 506 | * even ksmd can fail in this way - though it's usually breaking ksm |
| 507 | * just to undo a merge it made a moment before, so unlikely to oom. |
| 508 | * |
| 509 | * That's a pity: we might therefore have more kernel pages allocated |
| 510 | * than we're counting as nodes in the stable tree; but ksm_do_scan |
| 511 | * will retry to break_cow on each pass, so should recover the page |
| 512 | * in due course. The important thing is to not let VM_MERGEABLE |
| 513 | * be cleared while any such pages might remain in the area. |
| 514 | */ |
| 515 | return (ret & VM_FAULT_OOM) ? -ENOMEM : 0; |
| 516 | } |
| 517 | |
| 518 | static bool vma_ksm_compatible(struct vm_area_struct *vma) |
| 519 | { |
| 520 | if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE | VM_PFNMAP | |
| 521 | VM_IO | VM_DONTEXPAND | VM_HUGETLB | |
| 522 | VM_MIXEDMAP)) |
| 523 | return false; /* just ignore the advice */ |
| 524 | |
| 525 | if (vma_is_dax(vma)) |
| 526 | return false; |
| 527 | |
| 528 | #ifdef VM_SAO |
| 529 | if (vma->vm_flags & VM_SAO) |
| 530 | return false; |
| 531 | #endif |
| 532 | #ifdef VM_SPARC_ADI |
| 533 | if (vma->vm_flags & VM_SPARC_ADI) |
| 534 | return false; |
| 535 | #endif |
| 536 | |
| 537 | return true; |
| 538 | } |
| 539 | |
| 540 | static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm, |
| 541 | unsigned long addr) |
| 542 | { |
| 543 | struct vm_area_struct *vma; |
| 544 | if (ksm_test_exit(mm)) |
| 545 | return NULL; |
| 546 | vma = vma_lookup(mm, addr); |
| 547 | if (!vma || !(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) |
| 548 | return NULL; |
| 549 | return vma; |
| 550 | } |
| 551 | |
| 552 | static void break_cow(struct ksm_rmap_item *rmap_item) |
| 553 | { |
| 554 | struct mm_struct *mm = rmap_item->mm; |
| 555 | unsigned long addr = rmap_item->address; |
| 556 | struct vm_area_struct *vma; |
| 557 | |
| 558 | /* |
| 559 | * It is not an accident that whenever we want to break COW |
| 560 | * to undo, we also need to drop a reference to the anon_vma. |
| 561 | */ |
| 562 | put_anon_vma(rmap_item->anon_vma); |
| 563 | |
| 564 | mmap_read_lock(mm); |
| 565 | vma = find_mergeable_vma(mm, addr); |
| 566 | if (vma) |
| 567 | break_ksm(vma, addr); |
| 568 | mmap_read_unlock(mm); |
| 569 | } |
| 570 | |
| 571 | static struct page *get_mergeable_page(struct ksm_rmap_item *rmap_item) |
| 572 | { |
| 573 | struct mm_struct *mm = rmap_item->mm; |
| 574 | unsigned long addr = rmap_item->address; |
| 575 | struct vm_area_struct *vma; |
| 576 | struct page *page; |
| 577 | |
| 578 | mmap_read_lock(mm); |
| 579 | vma = find_mergeable_vma(mm, addr); |
| 580 | if (!vma) |
| 581 | goto out; |
| 582 | |
| 583 | page = follow_page(vma, addr, FOLL_GET); |
| 584 | if (IS_ERR_OR_NULL(page)) |
| 585 | goto out; |
| 586 | if (is_zone_device_page(page)) |
| 587 | goto out_putpage; |
| 588 | if (PageAnon(page)) { |
| 589 | flush_anon_page(vma, page, addr); |
| 590 | flush_dcache_page(page); |
| 591 | } else { |
| 592 | out_putpage: |
| 593 | put_page(page); |
| 594 | out: |
| 595 | page = NULL; |
| 596 | } |
| 597 | mmap_read_unlock(mm); |
| 598 | return page; |
| 599 | } |
| 600 | |
| 601 | /* |
| 602 | * This helper is used for getting right index into array of tree roots. |
| 603 | * When merge_across_nodes knob is set to 1, there are only two rb-trees for |
| 604 | * stable and unstable pages from all nodes with roots in index 0. Otherwise, |
| 605 | * every node has its own stable and unstable tree. |
| 606 | */ |
| 607 | static inline int get_kpfn_nid(unsigned long kpfn) |
| 608 | { |
| 609 | return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn)); |
| 610 | } |
| 611 | |
| 612 | static struct ksm_stable_node *alloc_stable_node_chain(struct ksm_stable_node *dup, |
| 613 | struct rb_root *root) |
| 614 | { |
| 615 | struct ksm_stable_node *chain = alloc_stable_node(); |
| 616 | VM_BUG_ON(is_stable_node_chain(dup)); |
| 617 | if (likely(chain)) { |
| 618 | INIT_HLIST_HEAD(&chain->hlist); |
| 619 | chain->chain_prune_time = jiffies; |
| 620 | chain->rmap_hlist_len = STABLE_NODE_CHAIN; |
| 621 | #if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA) |
| 622 | chain->nid = NUMA_NO_NODE; /* debug */ |
| 623 | #endif |
| 624 | ksm_stable_node_chains++; |
| 625 | |
| 626 | /* |
| 627 | * Put the stable node chain in the first dimension of |
| 628 | * the stable tree and at the same time remove the old |
| 629 | * stable node. |
| 630 | */ |
| 631 | rb_replace_node(&dup->node, &chain->node, root); |
| 632 | |
| 633 | /* |
| 634 | * Move the old stable node to the second dimension |
| 635 | * queued in the hlist_dup. The invariant is that all |
| 636 | * dup stable_nodes in the chain->hlist point to pages |
| 637 | * that are write protected and have the exact same |
| 638 | * content. |
| 639 | */ |
| 640 | stable_node_chain_add_dup(dup, chain); |
| 641 | } |
| 642 | return chain; |
| 643 | } |
| 644 | |
| 645 | static inline void free_stable_node_chain(struct ksm_stable_node *chain, |
| 646 | struct rb_root *root) |
| 647 | { |
| 648 | rb_erase(&chain->node, root); |
| 649 | free_stable_node(chain); |
| 650 | ksm_stable_node_chains--; |
| 651 | } |
| 652 | |
| 653 | static void remove_node_from_stable_tree(struct ksm_stable_node *stable_node) |
| 654 | { |
| 655 | struct ksm_rmap_item *rmap_item; |
| 656 | |
| 657 | /* check it's not STABLE_NODE_CHAIN or negative */ |
| 658 | BUG_ON(stable_node->rmap_hlist_len < 0); |
| 659 | |
| 660 | hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) { |
| 661 | if (rmap_item->hlist.next) { |
| 662 | ksm_pages_sharing--; |
| 663 | trace_ksm_remove_rmap_item(stable_node->kpfn, rmap_item, rmap_item->mm); |
| 664 | } else { |
| 665 | ksm_pages_shared--; |
| 666 | } |
| 667 | |
| 668 | rmap_item->mm->ksm_merging_pages--; |
| 669 | |
| 670 | VM_BUG_ON(stable_node->rmap_hlist_len <= 0); |
| 671 | stable_node->rmap_hlist_len--; |
| 672 | put_anon_vma(rmap_item->anon_vma); |
| 673 | rmap_item->address &= PAGE_MASK; |
| 674 | cond_resched(); |
| 675 | } |
| 676 | |
| 677 | /* |
| 678 | * We need the second aligned pointer of the migrate_nodes |
| 679 | * list_head to stay clear from the rb_parent_color union |
| 680 | * (aligned and different than any node) and also different |
| 681 | * from &migrate_nodes. This will verify that future list.h changes |
| 682 | * don't break STABLE_NODE_DUP_HEAD. Only recent gcc can handle it. |
| 683 | */ |
| 684 | BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes); |
| 685 | BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1); |
| 686 | |
| 687 | trace_ksm_remove_ksm_page(stable_node->kpfn); |
| 688 | if (stable_node->head == &migrate_nodes) |
| 689 | list_del(&stable_node->list); |
| 690 | else |
| 691 | stable_node_dup_del(stable_node); |
| 692 | free_stable_node(stable_node); |
| 693 | } |
| 694 | |
| 695 | enum get_ksm_page_flags { |
| 696 | GET_KSM_PAGE_NOLOCK, |
| 697 | GET_KSM_PAGE_LOCK, |
| 698 | GET_KSM_PAGE_TRYLOCK |
| 699 | }; |
| 700 | |
| 701 | /* |
| 702 | * get_ksm_page: checks if the page indicated by the stable node |
| 703 | * is still its ksm page, despite having held no reference to it. |
| 704 | * In which case we can trust the content of the page, and it |
| 705 | * returns the gotten page; but if the page has now been zapped, |
| 706 | * remove the stale node from the stable tree and return NULL. |
| 707 | * But beware, the stable node's page might be being migrated. |
| 708 | * |
| 709 | * You would expect the stable_node to hold a reference to the ksm page. |
| 710 | * But if it increments the page's count, swapping out has to wait for |
| 711 | * ksmd to come around again before it can free the page, which may take |
| 712 | * seconds or even minutes: much too unresponsive. So instead we use a |
| 713 | * "keyhole reference": access to the ksm page from the stable node peeps |
| 714 | * out through its keyhole to see if that page still holds the right key, |
| 715 | * pointing back to this stable node. This relies on freeing a PageAnon |
| 716 | * page to reset its page->mapping to NULL, and relies on no other use of |
| 717 | * a page to put something that might look like our key in page->mapping. |
| 718 | * is on its way to being freed; but it is an anomaly to bear in mind. |
| 719 | */ |
| 720 | static struct page *get_ksm_page(struct ksm_stable_node *stable_node, |
| 721 | enum get_ksm_page_flags flags) |
| 722 | { |
| 723 | struct page *page; |
| 724 | void *expected_mapping; |
| 725 | unsigned long kpfn; |
| 726 | |
| 727 | expected_mapping = (void *)((unsigned long)stable_node | |
| 728 | PAGE_MAPPING_KSM); |
| 729 | again: |
| 730 | kpfn = READ_ONCE(stable_node->kpfn); /* Address dependency. */ |
| 731 | page = pfn_to_page(kpfn); |
| 732 | if (READ_ONCE(page->mapping) != expected_mapping) |
| 733 | goto stale; |
| 734 | |
| 735 | /* |
| 736 | * We cannot do anything with the page while its refcount is 0. |
| 737 | * Usually 0 means free, or tail of a higher-order page: in which |
| 738 | * case this node is no longer referenced, and should be freed; |
| 739 | * however, it might mean that the page is under page_ref_freeze(). |
| 740 | * The __remove_mapping() case is easy, again the node is now stale; |
| 741 | * the same is in reuse_ksm_page() case; but if page is swapcache |
| 742 | * in folio_migrate_mapping(), it might still be our page, |
| 743 | * in which case it's essential to keep the node. |
| 744 | */ |
| 745 | while (!get_page_unless_zero(page)) { |
| 746 | /* |
| 747 | * Another check for page->mapping != expected_mapping would |
| 748 | * work here too. We have chosen the !PageSwapCache test to |
| 749 | * optimize the common case, when the page is or is about to |
| 750 | * be freed: PageSwapCache is cleared (under spin_lock_irq) |
| 751 | * in the ref_freeze section of __remove_mapping(); but Anon |
| 752 | * page->mapping reset to NULL later, in free_pages_prepare(). |
| 753 | */ |
| 754 | if (!PageSwapCache(page)) |
| 755 | goto stale; |
| 756 | cpu_relax(); |
| 757 | } |
| 758 | |
| 759 | if (READ_ONCE(page->mapping) != expected_mapping) { |
| 760 | put_page(page); |
| 761 | goto stale; |
| 762 | } |
| 763 | |
| 764 | if (flags == GET_KSM_PAGE_TRYLOCK) { |
| 765 | if (!trylock_page(page)) { |
| 766 | put_page(page); |
| 767 | return ERR_PTR(-EBUSY); |
| 768 | } |
| 769 | } else if (flags == GET_KSM_PAGE_LOCK) |
| 770 | lock_page(page); |
| 771 | |
| 772 | if (flags != GET_KSM_PAGE_NOLOCK) { |
| 773 | if (READ_ONCE(page->mapping) != expected_mapping) { |
| 774 | unlock_page(page); |
| 775 | put_page(page); |
| 776 | goto stale; |
| 777 | } |
| 778 | } |
| 779 | return page; |
| 780 | |
| 781 | stale: |
| 782 | /* |
| 783 | * We come here from above when page->mapping or !PageSwapCache |
| 784 | * suggests that the node is stale; but it might be under migration. |
| 785 | * We need smp_rmb(), matching the smp_wmb() in folio_migrate_ksm(), |
| 786 | * before checking whether node->kpfn has been changed. |
| 787 | */ |
| 788 | smp_rmb(); |
| 789 | if (READ_ONCE(stable_node->kpfn) != kpfn) |
| 790 | goto again; |
| 791 | remove_node_from_stable_tree(stable_node); |
| 792 | return NULL; |
| 793 | } |
| 794 | |
| 795 | /* |
| 796 | * Removing rmap_item from stable or unstable tree. |
| 797 | * This function will clean the information from the stable/unstable tree. |
| 798 | */ |
| 799 | static void remove_rmap_item_from_tree(struct ksm_rmap_item *rmap_item) |
| 800 | { |
| 801 | if (rmap_item->address & STABLE_FLAG) { |
| 802 | struct ksm_stable_node *stable_node; |
| 803 | struct page *page; |
| 804 | |
| 805 | stable_node = rmap_item->head; |
| 806 | page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK); |
| 807 | if (!page) |
| 808 | goto out; |
| 809 | |
| 810 | hlist_del(&rmap_item->hlist); |
| 811 | unlock_page(page); |
| 812 | put_page(page); |
| 813 | |
| 814 | if (!hlist_empty(&stable_node->hlist)) |
| 815 | ksm_pages_sharing--; |
| 816 | else |
| 817 | ksm_pages_shared--; |
| 818 | |
| 819 | rmap_item->mm->ksm_merging_pages--; |
| 820 | |
| 821 | VM_BUG_ON(stable_node->rmap_hlist_len <= 0); |
| 822 | stable_node->rmap_hlist_len--; |
| 823 | |
| 824 | put_anon_vma(rmap_item->anon_vma); |
| 825 | rmap_item->head = NULL; |
| 826 | rmap_item->address &= PAGE_MASK; |
| 827 | |
| 828 | } else if (rmap_item->address & UNSTABLE_FLAG) { |
| 829 | unsigned char age; |
| 830 | /* |
| 831 | * Usually ksmd can and must skip the rb_erase, because |
| 832 | * root_unstable_tree was already reset to RB_ROOT. |
| 833 | * But be careful when an mm is exiting: do the rb_erase |
| 834 | * if this rmap_item was inserted by this scan, rather |
| 835 | * than left over from before. |
| 836 | */ |
| 837 | age = (unsigned char)(ksm_scan.seqnr - rmap_item->address); |
| 838 | BUG_ON(age > 1); |
| 839 | if (!age) |
| 840 | rb_erase(&rmap_item->node, |
| 841 | root_unstable_tree + NUMA(rmap_item->nid)); |
| 842 | ksm_pages_unshared--; |
| 843 | rmap_item->address &= PAGE_MASK; |
| 844 | } |
| 845 | out: |
| 846 | cond_resched(); /* we're called from many long loops */ |
| 847 | } |
| 848 | |
| 849 | static void remove_trailing_rmap_items(struct ksm_rmap_item **rmap_list) |
| 850 | { |
| 851 | while (*rmap_list) { |
| 852 | struct ksm_rmap_item *rmap_item = *rmap_list; |
| 853 | *rmap_list = rmap_item->rmap_list; |
| 854 | remove_rmap_item_from_tree(rmap_item); |
| 855 | free_rmap_item(rmap_item); |
| 856 | } |
| 857 | } |
| 858 | |
| 859 | /* |
| 860 | * Though it's very tempting to unmerge rmap_items from stable tree rather |
| 861 | * than check every pte of a given vma, the locking doesn't quite work for |
| 862 | * that - an rmap_item is assigned to the stable tree after inserting ksm |
| 863 | * page and upping mmap_lock. Nor does it fit with the way we skip dup'ing |
| 864 | * rmap_items from parent to child at fork time (so as not to waste time |
| 865 | * if exit comes before the next scan reaches it). |
| 866 | * |
| 867 | * Similarly, although we'd like to remove rmap_items (so updating counts |
| 868 | * and freeing memory) when unmerging an area, it's easier to leave that |
| 869 | * to the next pass of ksmd - consider, for example, how ksmd might be |
| 870 | * in cmp_and_merge_page on one of the rmap_items we would be removing. |
| 871 | */ |
| 872 | static int unmerge_ksm_pages(struct vm_area_struct *vma, |
| 873 | unsigned long start, unsigned long end) |
| 874 | { |
| 875 | unsigned long addr; |
| 876 | int err = 0; |
| 877 | |
| 878 | for (addr = start; addr < end && !err; addr += PAGE_SIZE) { |
| 879 | if (ksm_test_exit(vma->vm_mm)) |
| 880 | break; |
| 881 | if (signal_pending(current)) |
| 882 | err = -ERESTARTSYS; |
| 883 | else |
| 884 | err = break_ksm(vma, addr); |
| 885 | } |
| 886 | return err; |
| 887 | } |
| 888 | |
| 889 | static inline struct ksm_stable_node *folio_stable_node(struct folio *folio) |
| 890 | { |
| 891 | return folio_test_ksm(folio) ? folio_raw_mapping(folio) : NULL; |
| 892 | } |
| 893 | |
| 894 | static inline struct ksm_stable_node *page_stable_node(struct page *page) |
| 895 | { |
| 896 | return folio_stable_node(page_folio(page)); |
| 897 | } |
| 898 | |
| 899 | static inline void set_page_stable_node(struct page *page, |
| 900 | struct ksm_stable_node *stable_node) |
| 901 | { |
| 902 | VM_BUG_ON_PAGE(PageAnon(page) && PageAnonExclusive(page), page); |
| 903 | page->mapping = (void *)((unsigned long)stable_node | PAGE_MAPPING_KSM); |
| 904 | } |
| 905 | |
| 906 | #ifdef CONFIG_SYSFS |
| 907 | /* |
| 908 | * Only called through the sysfs control interface: |
| 909 | */ |
| 910 | static int remove_stable_node(struct ksm_stable_node *stable_node) |
| 911 | { |
| 912 | struct page *page; |
| 913 | int err; |
| 914 | |
| 915 | page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK); |
| 916 | if (!page) { |
| 917 | /* |
| 918 | * get_ksm_page did remove_node_from_stable_tree itself. |
| 919 | */ |
| 920 | return 0; |
| 921 | } |
| 922 | |
| 923 | /* |
| 924 | * Page could be still mapped if this races with __mmput() running in |
| 925 | * between ksm_exit() and exit_mmap(). Just refuse to let |
| 926 | * merge_across_nodes/max_page_sharing be switched. |
| 927 | */ |
| 928 | err = -EBUSY; |
| 929 | if (!page_mapped(page)) { |
| 930 | /* |
| 931 | * The stable node did not yet appear stale to get_ksm_page(), |
| 932 | * since that allows for an unmapped ksm page to be recognized |
| 933 | * right up until it is freed; but the node is safe to remove. |
| 934 | * This page might be in a pagevec waiting to be freed, |
| 935 | * or it might be PageSwapCache (perhaps under writeback), |
| 936 | * or it might have been removed from swapcache a moment ago. |
| 937 | */ |
| 938 | set_page_stable_node(page, NULL); |
| 939 | remove_node_from_stable_tree(stable_node); |
| 940 | err = 0; |
| 941 | } |
| 942 | |
| 943 | unlock_page(page); |
| 944 | put_page(page); |
| 945 | return err; |
| 946 | } |
| 947 | |
| 948 | static int remove_stable_node_chain(struct ksm_stable_node *stable_node, |
| 949 | struct rb_root *root) |
| 950 | { |
| 951 | struct ksm_stable_node *dup; |
| 952 | struct hlist_node *hlist_safe; |
| 953 | |
| 954 | if (!is_stable_node_chain(stable_node)) { |
| 955 | VM_BUG_ON(is_stable_node_dup(stable_node)); |
| 956 | if (remove_stable_node(stable_node)) |
| 957 | return true; |
| 958 | else |
| 959 | return false; |
| 960 | } |
| 961 | |
| 962 | hlist_for_each_entry_safe(dup, hlist_safe, |
| 963 | &stable_node->hlist, hlist_dup) { |
| 964 | VM_BUG_ON(!is_stable_node_dup(dup)); |
| 965 | if (remove_stable_node(dup)) |
| 966 | return true; |
| 967 | } |
| 968 | BUG_ON(!hlist_empty(&stable_node->hlist)); |
| 969 | free_stable_node_chain(stable_node, root); |
| 970 | return false; |
| 971 | } |
| 972 | |
| 973 | static int remove_all_stable_nodes(void) |
| 974 | { |
| 975 | struct ksm_stable_node *stable_node, *next; |
| 976 | int nid; |
| 977 | int err = 0; |
| 978 | |
| 979 | for (nid = 0; nid < ksm_nr_node_ids; nid++) { |
| 980 | while (root_stable_tree[nid].rb_node) { |
| 981 | stable_node = rb_entry(root_stable_tree[nid].rb_node, |
| 982 | struct ksm_stable_node, node); |
| 983 | if (remove_stable_node_chain(stable_node, |
| 984 | root_stable_tree + nid)) { |
| 985 | err = -EBUSY; |
| 986 | break; /* proceed to next nid */ |
| 987 | } |
| 988 | cond_resched(); |
| 989 | } |
| 990 | } |
| 991 | list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) { |
| 992 | if (remove_stable_node(stable_node)) |
| 993 | err = -EBUSY; |
| 994 | cond_resched(); |
| 995 | } |
| 996 | return err; |
| 997 | } |
| 998 | |
| 999 | static int unmerge_and_remove_all_rmap_items(void) |
| 1000 | { |
| 1001 | struct ksm_mm_slot *mm_slot; |
| 1002 | struct mm_slot *slot; |
| 1003 | struct mm_struct *mm; |
| 1004 | struct vm_area_struct *vma; |
| 1005 | int err = 0; |
| 1006 | |
| 1007 | spin_lock(&ksm_mmlist_lock); |
| 1008 | slot = list_entry(ksm_mm_head.slot.mm_node.next, |
| 1009 | struct mm_slot, mm_node); |
| 1010 | ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); |
| 1011 | spin_unlock(&ksm_mmlist_lock); |
| 1012 | |
| 1013 | for (mm_slot = ksm_scan.mm_slot; mm_slot != &ksm_mm_head; |
| 1014 | mm_slot = ksm_scan.mm_slot) { |
| 1015 | VMA_ITERATOR(vmi, mm_slot->slot.mm, 0); |
| 1016 | |
| 1017 | mm = mm_slot->slot.mm; |
| 1018 | mmap_read_lock(mm); |
| 1019 | |
| 1020 | /* |
| 1021 | * Exit right away if mm is exiting to avoid lockdep issue in |
| 1022 | * the maple tree |
| 1023 | */ |
| 1024 | if (ksm_test_exit(mm)) |
| 1025 | goto mm_exiting; |
| 1026 | |
| 1027 | for_each_vma(vmi, vma) { |
| 1028 | if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) |
| 1029 | continue; |
| 1030 | err = unmerge_ksm_pages(vma, |
| 1031 | vma->vm_start, vma->vm_end); |
| 1032 | if (err) |
| 1033 | goto error; |
| 1034 | } |
| 1035 | |
| 1036 | mm_exiting: |
| 1037 | remove_trailing_rmap_items(&mm_slot->rmap_list); |
| 1038 | mmap_read_unlock(mm); |
| 1039 | |
| 1040 | spin_lock(&ksm_mmlist_lock); |
| 1041 | slot = list_entry(mm_slot->slot.mm_node.next, |
| 1042 | struct mm_slot, mm_node); |
| 1043 | ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); |
| 1044 | if (ksm_test_exit(mm)) { |
| 1045 | hash_del(&mm_slot->slot.hash); |
| 1046 | list_del(&mm_slot->slot.mm_node); |
| 1047 | spin_unlock(&ksm_mmlist_lock); |
| 1048 | |
| 1049 | mm_slot_free(mm_slot_cache, mm_slot); |
| 1050 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); |
| 1051 | clear_bit(MMF_VM_MERGE_ANY, &mm->flags); |
| 1052 | mmdrop(mm); |
| 1053 | } else |
| 1054 | spin_unlock(&ksm_mmlist_lock); |
| 1055 | } |
| 1056 | |
| 1057 | /* Clean up stable nodes, but don't worry if some are still busy */ |
| 1058 | remove_all_stable_nodes(); |
| 1059 | ksm_scan.seqnr = 0; |
| 1060 | return 0; |
| 1061 | |
| 1062 | error: |
| 1063 | mmap_read_unlock(mm); |
| 1064 | spin_lock(&ksm_mmlist_lock); |
| 1065 | ksm_scan.mm_slot = &ksm_mm_head; |
| 1066 | spin_unlock(&ksm_mmlist_lock); |
| 1067 | return err; |
| 1068 | } |
| 1069 | #endif /* CONFIG_SYSFS */ |
| 1070 | |
| 1071 | static u32 calc_checksum(struct page *page) |
| 1072 | { |
| 1073 | u32 checksum; |
| 1074 | void *addr = kmap_atomic(page); |
| 1075 | checksum = xxhash(addr, PAGE_SIZE, 0); |
| 1076 | kunmap_atomic(addr); |
| 1077 | return checksum; |
| 1078 | } |
| 1079 | |
| 1080 | static int write_protect_page(struct vm_area_struct *vma, struct page *page, |
| 1081 | pte_t *orig_pte) |
| 1082 | { |
| 1083 | struct mm_struct *mm = vma->vm_mm; |
| 1084 | DEFINE_PAGE_VMA_WALK(pvmw, page, vma, 0, 0); |
| 1085 | int swapped; |
| 1086 | int err = -EFAULT; |
| 1087 | struct mmu_notifier_range range; |
| 1088 | bool anon_exclusive; |
| 1089 | |
| 1090 | pvmw.address = page_address_in_vma(page, vma); |
| 1091 | if (pvmw.address == -EFAULT) |
| 1092 | goto out; |
| 1093 | |
| 1094 | BUG_ON(PageTransCompound(page)); |
| 1095 | |
| 1096 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, pvmw.address, |
| 1097 | pvmw.address + PAGE_SIZE); |
| 1098 | mmu_notifier_invalidate_range_start(&range); |
| 1099 | |
| 1100 | if (!page_vma_mapped_walk(&pvmw)) |
| 1101 | goto out_mn; |
| 1102 | if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?")) |
| 1103 | goto out_unlock; |
| 1104 | |
| 1105 | anon_exclusive = PageAnonExclusive(page); |
| 1106 | if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) || |
| 1107 | anon_exclusive || mm_tlb_flush_pending(mm)) { |
| 1108 | pte_t entry; |
| 1109 | |
| 1110 | swapped = PageSwapCache(page); |
| 1111 | flush_cache_page(vma, pvmw.address, page_to_pfn(page)); |
| 1112 | /* |
| 1113 | * Ok this is tricky, when get_user_pages_fast() run it doesn't |
| 1114 | * take any lock, therefore the check that we are going to make |
| 1115 | * with the pagecount against the mapcount is racy and |
| 1116 | * O_DIRECT can happen right after the check. |
| 1117 | * So we clear the pte and flush the tlb before the check |
| 1118 | * this assure us that no O_DIRECT can happen after the check |
| 1119 | * or in the middle of the check. |
| 1120 | * |
| 1121 | * No need to notify as we are downgrading page table to read |
| 1122 | * only not changing it to point to a new page. |
| 1123 | * |
| 1124 | * See Documentation/mm/mmu_notifier.rst |
| 1125 | */ |
| 1126 | entry = ptep_clear_flush(vma, pvmw.address, pvmw.pte); |
| 1127 | /* |
| 1128 | * Check that no O_DIRECT or similar I/O is in progress on the |
| 1129 | * page |
| 1130 | */ |
| 1131 | if (page_mapcount(page) + 1 + swapped != page_count(page)) { |
| 1132 | set_pte_at(mm, pvmw.address, pvmw.pte, entry); |
| 1133 | goto out_unlock; |
| 1134 | } |
| 1135 | |
| 1136 | /* See page_try_share_anon_rmap(): clear PTE first. */ |
| 1137 | if (anon_exclusive && page_try_share_anon_rmap(page)) { |
| 1138 | set_pte_at(mm, pvmw.address, pvmw.pte, entry); |
| 1139 | goto out_unlock; |
| 1140 | } |
| 1141 | |
| 1142 | if (pte_dirty(entry)) |
| 1143 | set_page_dirty(page); |
| 1144 | entry = pte_mkclean(entry); |
| 1145 | |
| 1146 | if (pte_write(entry)) |
| 1147 | entry = pte_wrprotect(entry); |
| 1148 | |
| 1149 | set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry); |
| 1150 | } |
| 1151 | *orig_pte = *pvmw.pte; |
| 1152 | err = 0; |
| 1153 | |
| 1154 | out_unlock: |
| 1155 | page_vma_mapped_walk_done(&pvmw); |
| 1156 | out_mn: |
| 1157 | mmu_notifier_invalidate_range_end(&range); |
| 1158 | out: |
| 1159 | return err; |
| 1160 | } |
| 1161 | |
| 1162 | /** |
| 1163 | * replace_page - replace page in vma by new ksm page |
| 1164 | * @vma: vma that holds the pte pointing to page |
| 1165 | * @page: the page we are replacing by kpage |
| 1166 | * @kpage: the ksm page we replace page by |
| 1167 | * @orig_pte: the original value of the pte |
| 1168 | * |
| 1169 | * Returns 0 on success, -EFAULT on failure. |
| 1170 | */ |
| 1171 | static int replace_page(struct vm_area_struct *vma, struct page *page, |
| 1172 | struct page *kpage, pte_t orig_pte) |
| 1173 | { |
| 1174 | struct mm_struct *mm = vma->vm_mm; |
| 1175 | struct folio *folio; |
| 1176 | pmd_t *pmd; |
| 1177 | pmd_t pmde; |
| 1178 | pte_t *ptep; |
| 1179 | pte_t newpte; |
| 1180 | spinlock_t *ptl; |
| 1181 | unsigned long addr; |
| 1182 | int err = -EFAULT; |
| 1183 | struct mmu_notifier_range range; |
| 1184 | |
| 1185 | addr = page_address_in_vma(page, vma); |
| 1186 | if (addr == -EFAULT) |
| 1187 | goto out; |
| 1188 | |
| 1189 | pmd = mm_find_pmd(mm, addr); |
| 1190 | if (!pmd) |
| 1191 | goto out; |
| 1192 | /* |
| 1193 | * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at() |
| 1194 | * without holding anon_vma lock for write. So when looking for a |
| 1195 | * genuine pmde (in which to find pte), test present and !THP together. |
| 1196 | */ |
| 1197 | pmde = *pmd; |
| 1198 | barrier(); |
| 1199 | if (!pmd_present(pmde) || pmd_trans_huge(pmde)) |
| 1200 | goto out; |
| 1201 | |
| 1202 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, addr, |
| 1203 | addr + PAGE_SIZE); |
| 1204 | mmu_notifier_invalidate_range_start(&range); |
| 1205 | |
| 1206 | ptep = pte_offset_map_lock(mm, pmd, addr, &ptl); |
| 1207 | if (!pte_same(*ptep, orig_pte)) { |
| 1208 | pte_unmap_unlock(ptep, ptl); |
| 1209 | goto out_mn; |
| 1210 | } |
| 1211 | VM_BUG_ON_PAGE(PageAnonExclusive(page), page); |
| 1212 | VM_BUG_ON_PAGE(PageAnon(kpage) && PageAnonExclusive(kpage), kpage); |
| 1213 | |
| 1214 | /* |
| 1215 | * No need to check ksm_use_zero_pages here: we can only have a |
| 1216 | * zero_page here if ksm_use_zero_pages was enabled already. |
| 1217 | */ |
| 1218 | if (!is_zero_pfn(page_to_pfn(kpage))) { |
| 1219 | get_page(kpage); |
| 1220 | page_add_anon_rmap(kpage, vma, addr, RMAP_NONE); |
| 1221 | newpte = mk_pte(kpage, vma->vm_page_prot); |
| 1222 | } else { |
| 1223 | newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage), |
| 1224 | vma->vm_page_prot)); |
| 1225 | /* |
| 1226 | * We're replacing an anonymous page with a zero page, which is |
| 1227 | * not anonymous. We need to do proper accounting otherwise we |
| 1228 | * will get wrong values in /proc, and a BUG message in dmesg |
| 1229 | * when tearing down the mm. |
| 1230 | */ |
| 1231 | dec_mm_counter(mm, MM_ANONPAGES); |
| 1232 | } |
| 1233 | |
| 1234 | flush_cache_page(vma, addr, pte_pfn(*ptep)); |
| 1235 | /* |
| 1236 | * No need to notify as we are replacing a read only page with another |
| 1237 | * read only page with the same content. |
| 1238 | * |
| 1239 | * See Documentation/mm/mmu_notifier.rst |
| 1240 | */ |
| 1241 | ptep_clear_flush(vma, addr, ptep); |
| 1242 | set_pte_at_notify(mm, addr, ptep, newpte); |
| 1243 | |
| 1244 | folio = page_folio(page); |
| 1245 | page_remove_rmap(page, vma, false); |
| 1246 | if (!folio_mapped(folio)) |
| 1247 | folio_free_swap(folio); |
| 1248 | folio_put(folio); |
| 1249 | |
| 1250 | pte_unmap_unlock(ptep, ptl); |
| 1251 | err = 0; |
| 1252 | out_mn: |
| 1253 | mmu_notifier_invalidate_range_end(&range); |
| 1254 | out: |
| 1255 | return err; |
| 1256 | } |
| 1257 | |
| 1258 | /* |
| 1259 | * try_to_merge_one_page - take two pages and merge them into one |
| 1260 | * @vma: the vma that holds the pte pointing to page |
| 1261 | * @page: the PageAnon page that we want to replace with kpage |
| 1262 | * @kpage: the PageKsm page that we want to map instead of page, |
| 1263 | * or NULL the first time when we want to use page as kpage. |
| 1264 | * |
| 1265 | * This function returns 0 if the pages were merged, -EFAULT otherwise. |
| 1266 | */ |
| 1267 | static int try_to_merge_one_page(struct vm_area_struct *vma, |
| 1268 | struct page *page, struct page *kpage) |
| 1269 | { |
| 1270 | pte_t orig_pte = __pte(0); |
| 1271 | int err = -EFAULT; |
| 1272 | |
| 1273 | if (page == kpage) /* ksm page forked */ |
| 1274 | return 0; |
| 1275 | |
| 1276 | if (!PageAnon(page)) |
| 1277 | goto out; |
| 1278 | |
| 1279 | /* |
| 1280 | * We need the page lock to read a stable PageSwapCache in |
| 1281 | * write_protect_page(). We use trylock_page() instead of |
| 1282 | * lock_page() because we don't want to wait here - we |
| 1283 | * prefer to continue scanning and merging different pages, |
| 1284 | * then come back to this page when it is unlocked. |
| 1285 | */ |
| 1286 | if (!trylock_page(page)) |
| 1287 | goto out; |
| 1288 | |
| 1289 | if (PageTransCompound(page)) { |
| 1290 | if (split_huge_page(page)) |
| 1291 | goto out_unlock; |
| 1292 | } |
| 1293 | |
| 1294 | /* |
| 1295 | * If this anonymous page is mapped only here, its pte may need |
| 1296 | * to be write-protected. If it's mapped elsewhere, all of its |
| 1297 | * ptes are necessarily already write-protected. But in either |
| 1298 | * case, we need to lock and check page_count is not raised. |
| 1299 | */ |
| 1300 | if (write_protect_page(vma, page, &orig_pte) == 0) { |
| 1301 | if (!kpage) { |
| 1302 | /* |
| 1303 | * While we hold page lock, upgrade page from |
| 1304 | * PageAnon+anon_vma to PageKsm+NULL stable_node: |
| 1305 | * stable_tree_insert() will update stable_node. |
| 1306 | */ |
| 1307 | set_page_stable_node(page, NULL); |
| 1308 | mark_page_accessed(page); |
| 1309 | /* |
| 1310 | * Page reclaim just frees a clean page with no dirty |
| 1311 | * ptes: make sure that the ksm page would be swapped. |
| 1312 | */ |
| 1313 | if (!PageDirty(page)) |
| 1314 | SetPageDirty(page); |
| 1315 | err = 0; |
| 1316 | } else if (pages_identical(page, kpage)) |
| 1317 | err = replace_page(vma, page, kpage, orig_pte); |
| 1318 | } |
| 1319 | |
| 1320 | out_unlock: |
| 1321 | unlock_page(page); |
| 1322 | out: |
| 1323 | return err; |
| 1324 | } |
| 1325 | |
| 1326 | /* |
| 1327 | * try_to_merge_with_ksm_page - like try_to_merge_two_pages, |
| 1328 | * but no new kernel page is allocated: kpage must already be a ksm page. |
| 1329 | * |
| 1330 | * This function returns 0 if the pages were merged, -EFAULT otherwise. |
| 1331 | */ |
| 1332 | static int try_to_merge_with_ksm_page(struct ksm_rmap_item *rmap_item, |
| 1333 | struct page *page, struct page *kpage) |
| 1334 | { |
| 1335 | struct mm_struct *mm = rmap_item->mm; |
| 1336 | struct vm_area_struct *vma; |
| 1337 | int err = -EFAULT; |
| 1338 | |
| 1339 | mmap_read_lock(mm); |
| 1340 | vma = find_mergeable_vma(mm, rmap_item->address); |
| 1341 | if (!vma) |
| 1342 | goto out; |
| 1343 | |
| 1344 | err = try_to_merge_one_page(vma, page, kpage); |
| 1345 | if (err) |
| 1346 | goto out; |
| 1347 | |
| 1348 | /* Unstable nid is in union with stable anon_vma: remove first */ |
| 1349 | remove_rmap_item_from_tree(rmap_item); |
| 1350 | |
| 1351 | /* Must get reference to anon_vma while still holding mmap_lock */ |
| 1352 | rmap_item->anon_vma = vma->anon_vma; |
| 1353 | get_anon_vma(vma->anon_vma); |
| 1354 | out: |
| 1355 | mmap_read_unlock(mm); |
| 1356 | trace_ksm_merge_with_ksm_page(kpage, page_to_pfn(kpage ? kpage : page), |
| 1357 | rmap_item, mm, err); |
| 1358 | return err; |
| 1359 | } |
| 1360 | |
| 1361 | /* |
| 1362 | * try_to_merge_two_pages - take two identical pages and prepare them |
| 1363 | * to be merged into one page. |
| 1364 | * |
| 1365 | * This function returns the kpage if we successfully merged two identical |
| 1366 | * pages into one ksm page, NULL otherwise. |
| 1367 | * |
| 1368 | * Note that this function upgrades page to ksm page: if one of the pages |
| 1369 | * is already a ksm page, try_to_merge_with_ksm_page should be used. |
| 1370 | */ |
| 1371 | static struct page *try_to_merge_two_pages(struct ksm_rmap_item *rmap_item, |
| 1372 | struct page *page, |
| 1373 | struct ksm_rmap_item *tree_rmap_item, |
| 1374 | struct page *tree_page) |
| 1375 | { |
| 1376 | int err; |
| 1377 | |
| 1378 | err = try_to_merge_with_ksm_page(rmap_item, page, NULL); |
| 1379 | if (!err) { |
| 1380 | err = try_to_merge_with_ksm_page(tree_rmap_item, |
| 1381 | tree_page, page); |
| 1382 | /* |
| 1383 | * If that fails, we have a ksm page with only one pte |
| 1384 | * pointing to it: so break it. |
| 1385 | */ |
| 1386 | if (err) |
| 1387 | break_cow(rmap_item); |
| 1388 | } |
| 1389 | return err ? NULL : page; |
| 1390 | } |
| 1391 | |
| 1392 | static __always_inline |
| 1393 | bool __is_page_sharing_candidate(struct ksm_stable_node *stable_node, int offset) |
| 1394 | { |
| 1395 | VM_BUG_ON(stable_node->rmap_hlist_len < 0); |
| 1396 | /* |
| 1397 | * Check that at least one mapping still exists, otherwise |
| 1398 | * there's no much point to merge and share with this |
| 1399 | * stable_node, as the underlying tree_page of the other |
| 1400 | * sharer is going to be freed soon. |
| 1401 | */ |
| 1402 | return stable_node->rmap_hlist_len && |
| 1403 | stable_node->rmap_hlist_len + offset < ksm_max_page_sharing; |
| 1404 | } |
| 1405 | |
| 1406 | static __always_inline |
| 1407 | bool is_page_sharing_candidate(struct ksm_stable_node *stable_node) |
| 1408 | { |
| 1409 | return __is_page_sharing_candidate(stable_node, 0); |
| 1410 | } |
| 1411 | |
| 1412 | static struct page *stable_node_dup(struct ksm_stable_node **_stable_node_dup, |
| 1413 | struct ksm_stable_node **_stable_node, |
| 1414 | struct rb_root *root, |
| 1415 | bool prune_stale_stable_nodes) |
| 1416 | { |
| 1417 | struct ksm_stable_node *dup, *found = NULL, *stable_node = *_stable_node; |
| 1418 | struct hlist_node *hlist_safe; |
| 1419 | struct page *_tree_page, *tree_page = NULL; |
| 1420 | int nr = 0; |
| 1421 | int found_rmap_hlist_len; |
| 1422 | |
| 1423 | if (!prune_stale_stable_nodes || |
| 1424 | time_before(jiffies, stable_node->chain_prune_time + |
| 1425 | msecs_to_jiffies( |
| 1426 | ksm_stable_node_chains_prune_millisecs))) |
| 1427 | prune_stale_stable_nodes = false; |
| 1428 | else |
| 1429 | stable_node->chain_prune_time = jiffies; |
| 1430 | |
| 1431 | hlist_for_each_entry_safe(dup, hlist_safe, |
| 1432 | &stable_node->hlist, hlist_dup) { |
| 1433 | cond_resched(); |
| 1434 | /* |
| 1435 | * We must walk all stable_node_dup to prune the stale |
| 1436 | * stable nodes during lookup. |
| 1437 | * |
| 1438 | * get_ksm_page can drop the nodes from the |
| 1439 | * stable_node->hlist if they point to freed pages |
| 1440 | * (that's why we do a _safe walk). The "dup" |
| 1441 | * stable_node parameter itself will be freed from |
| 1442 | * under us if it returns NULL. |
| 1443 | */ |
| 1444 | _tree_page = get_ksm_page(dup, GET_KSM_PAGE_NOLOCK); |
| 1445 | if (!_tree_page) |
| 1446 | continue; |
| 1447 | nr += 1; |
| 1448 | if (is_page_sharing_candidate(dup)) { |
| 1449 | if (!found || |
| 1450 | dup->rmap_hlist_len > found_rmap_hlist_len) { |
| 1451 | if (found) |
| 1452 | put_page(tree_page); |
| 1453 | found = dup; |
| 1454 | found_rmap_hlist_len = found->rmap_hlist_len; |
| 1455 | tree_page = _tree_page; |
| 1456 | |
| 1457 | /* skip put_page for found dup */ |
| 1458 | if (!prune_stale_stable_nodes) |
| 1459 | break; |
| 1460 | continue; |
| 1461 | } |
| 1462 | } |
| 1463 | put_page(_tree_page); |
| 1464 | } |
| 1465 | |
| 1466 | if (found) { |
| 1467 | /* |
| 1468 | * nr is counting all dups in the chain only if |
| 1469 | * prune_stale_stable_nodes is true, otherwise we may |
| 1470 | * break the loop at nr == 1 even if there are |
| 1471 | * multiple entries. |
| 1472 | */ |
| 1473 | if (prune_stale_stable_nodes && nr == 1) { |
| 1474 | /* |
| 1475 | * If there's not just one entry it would |
| 1476 | * corrupt memory, better BUG_ON. In KSM |
| 1477 | * context with no lock held it's not even |
| 1478 | * fatal. |
| 1479 | */ |
| 1480 | BUG_ON(stable_node->hlist.first->next); |
| 1481 | |
| 1482 | /* |
| 1483 | * There's just one entry and it is below the |
| 1484 | * deduplication limit so drop the chain. |
| 1485 | */ |
| 1486 | rb_replace_node(&stable_node->node, &found->node, |
| 1487 | root); |
| 1488 | free_stable_node(stable_node); |
| 1489 | ksm_stable_node_chains--; |
| 1490 | ksm_stable_node_dups--; |
| 1491 | /* |
| 1492 | * NOTE: the caller depends on the stable_node |
| 1493 | * to be equal to stable_node_dup if the chain |
| 1494 | * was collapsed. |
| 1495 | */ |
| 1496 | *_stable_node = found; |
| 1497 | /* |
| 1498 | * Just for robustness, as stable_node is |
| 1499 | * otherwise left as a stable pointer, the |
| 1500 | * compiler shall optimize it away at build |
| 1501 | * time. |
| 1502 | */ |
| 1503 | stable_node = NULL; |
| 1504 | } else if (stable_node->hlist.first != &found->hlist_dup && |
| 1505 | __is_page_sharing_candidate(found, 1)) { |
| 1506 | /* |
| 1507 | * If the found stable_node dup can accept one |
| 1508 | * more future merge (in addition to the one |
| 1509 | * that is underway) and is not at the head of |
| 1510 | * the chain, put it there so next search will |
| 1511 | * be quicker in the !prune_stale_stable_nodes |
| 1512 | * case. |
| 1513 | * |
| 1514 | * NOTE: it would be inaccurate to use nr > 1 |
| 1515 | * instead of checking the hlist.first pointer |
| 1516 | * directly, because in the |
| 1517 | * prune_stale_stable_nodes case "nr" isn't |
| 1518 | * the position of the found dup in the chain, |
| 1519 | * but the total number of dups in the chain. |
| 1520 | */ |
| 1521 | hlist_del(&found->hlist_dup); |
| 1522 | hlist_add_head(&found->hlist_dup, |
| 1523 | &stable_node->hlist); |
| 1524 | } |
| 1525 | } |
| 1526 | |
| 1527 | *_stable_node_dup = found; |
| 1528 | return tree_page; |
| 1529 | } |
| 1530 | |
| 1531 | static struct ksm_stable_node *stable_node_dup_any(struct ksm_stable_node *stable_node, |
| 1532 | struct rb_root *root) |
| 1533 | { |
| 1534 | if (!is_stable_node_chain(stable_node)) |
| 1535 | return stable_node; |
| 1536 | if (hlist_empty(&stable_node->hlist)) { |
| 1537 | free_stable_node_chain(stable_node, root); |
| 1538 | return NULL; |
| 1539 | } |
| 1540 | return hlist_entry(stable_node->hlist.first, |
| 1541 | typeof(*stable_node), hlist_dup); |
| 1542 | } |
| 1543 | |
| 1544 | /* |
| 1545 | * Like for get_ksm_page, this function can free the *_stable_node and |
| 1546 | * *_stable_node_dup if the returned tree_page is NULL. |
| 1547 | * |
| 1548 | * It can also free and overwrite *_stable_node with the found |
| 1549 | * stable_node_dup if the chain is collapsed (in which case |
| 1550 | * *_stable_node will be equal to *_stable_node_dup like if the chain |
| 1551 | * never existed). It's up to the caller to verify tree_page is not |
| 1552 | * NULL before dereferencing *_stable_node or *_stable_node_dup. |
| 1553 | * |
| 1554 | * *_stable_node_dup is really a second output parameter of this |
| 1555 | * function and will be overwritten in all cases, the caller doesn't |
| 1556 | * need to initialize it. |
| 1557 | */ |
| 1558 | static struct page *__stable_node_chain(struct ksm_stable_node **_stable_node_dup, |
| 1559 | struct ksm_stable_node **_stable_node, |
| 1560 | struct rb_root *root, |
| 1561 | bool prune_stale_stable_nodes) |
| 1562 | { |
| 1563 | struct ksm_stable_node *stable_node = *_stable_node; |
| 1564 | if (!is_stable_node_chain(stable_node)) { |
| 1565 | if (is_page_sharing_candidate(stable_node)) { |
| 1566 | *_stable_node_dup = stable_node; |
| 1567 | return get_ksm_page(stable_node, GET_KSM_PAGE_NOLOCK); |
| 1568 | } |
| 1569 | /* |
| 1570 | * _stable_node_dup set to NULL means the stable_node |
| 1571 | * reached the ksm_max_page_sharing limit. |
| 1572 | */ |
| 1573 | *_stable_node_dup = NULL; |
| 1574 | return NULL; |
| 1575 | } |
| 1576 | return stable_node_dup(_stable_node_dup, _stable_node, root, |
| 1577 | prune_stale_stable_nodes); |
| 1578 | } |
| 1579 | |
| 1580 | static __always_inline struct page *chain_prune(struct ksm_stable_node **s_n_d, |
| 1581 | struct ksm_stable_node **s_n, |
| 1582 | struct rb_root *root) |
| 1583 | { |
| 1584 | return __stable_node_chain(s_n_d, s_n, root, true); |
| 1585 | } |
| 1586 | |
| 1587 | static __always_inline struct page *chain(struct ksm_stable_node **s_n_d, |
| 1588 | struct ksm_stable_node *s_n, |
| 1589 | struct rb_root *root) |
| 1590 | { |
| 1591 | struct ksm_stable_node *old_stable_node = s_n; |
| 1592 | struct page *tree_page; |
| 1593 | |
| 1594 | tree_page = __stable_node_chain(s_n_d, &s_n, root, false); |
| 1595 | /* not pruning dups so s_n cannot have changed */ |
| 1596 | VM_BUG_ON(s_n != old_stable_node); |
| 1597 | return tree_page; |
| 1598 | } |
| 1599 | |
| 1600 | /* |
| 1601 | * stable_tree_search - search for page inside the stable tree |
| 1602 | * |
| 1603 | * This function checks if there is a page inside the stable tree |
| 1604 | * with identical content to the page that we are scanning right now. |
| 1605 | * |
| 1606 | * This function returns the stable tree node of identical content if found, |
| 1607 | * NULL otherwise. |
| 1608 | */ |
| 1609 | static struct page *stable_tree_search(struct page *page) |
| 1610 | { |
| 1611 | int nid; |
| 1612 | struct rb_root *root; |
| 1613 | struct rb_node **new; |
| 1614 | struct rb_node *parent; |
| 1615 | struct ksm_stable_node *stable_node, *stable_node_dup, *stable_node_any; |
| 1616 | struct ksm_stable_node *page_node; |
| 1617 | |
| 1618 | page_node = page_stable_node(page); |
| 1619 | if (page_node && page_node->head != &migrate_nodes) { |
| 1620 | /* ksm page forked */ |
| 1621 | get_page(page); |
| 1622 | return page; |
| 1623 | } |
| 1624 | |
| 1625 | nid = get_kpfn_nid(page_to_pfn(page)); |
| 1626 | root = root_stable_tree + nid; |
| 1627 | again: |
| 1628 | new = &root->rb_node; |
| 1629 | parent = NULL; |
| 1630 | |
| 1631 | while (*new) { |
| 1632 | struct page *tree_page; |
| 1633 | int ret; |
| 1634 | |
| 1635 | cond_resched(); |
| 1636 | stable_node = rb_entry(*new, struct ksm_stable_node, node); |
| 1637 | stable_node_any = NULL; |
| 1638 | tree_page = chain_prune(&stable_node_dup, &stable_node, root); |
| 1639 | /* |
| 1640 | * NOTE: stable_node may have been freed by |
| 1641 | * chain_prune() if the returned stable_node_dup is |
| 1642 | * not NULL. stable_node_dup may have been inserted in |
| 1643 | * the rbtree instead as a regular stable_node (in |
| 1644 | * order to collapse the stable_node chain if a single |
| 1645 | * stable_node dup was found in it). In such case the |
| 1646 | * stable_node is overwritten by the callee to point |
| 1647 | * to the stable_node_dup that was collapsed in the |
| 1648 | * stable rbtree and stable_node will be equal to |
| 1649 | * stable_node_dup like if the chain never existed. |
| 1650 | */ |
| 1651 | if (!stable_node_dup) { |
| 1652 | /* |
| 1653 | * Either all stable_node dups were full in |
| 1654 | * this stable_node chain, or this chain was |
| 1655 | * empty and should be rb_erased. |
| 1656 | */ |
| 1657 | stable_node_any = stable_node_dup_any(stable_node, |
| 1658 | root); |
| 1659 | if (!stable_node_any) { |
| 1660 | /* rb_erase just run */ |
| 1661 | goto again; |
| 1662 | } |
| 1663 | /* |
| 1664 | * Take any of the stable_node dups page of |
| 1665 | * this stable_node chain to let the tree walk |
| 1666 | * continue. All KSM pages belonging to the |
| 1667 | * stable_node dups in a stable_node chain |
| 1668 | * have the same content and they're |
| 1669 | * write protected at all times. Any will work |
| 1670 | * fine to continue the walk. |
| 1671 | */ |
| 1672 | tree_page = get_ksm_page(stable_node_any, |
| 1673 | GET_KSM_PAGE_NOLOCK); |
| 1674 | } |
| 1675 | VM_BUG_ON(!stable_node_dup ^ !!stable_node_any); |
| 1676 | if (!tree_page) { |
| 1677 | /* |
| 1678 | * If we walked over a stale stable_node, |
| 1679 | * get_ksm_page() will call rb_erase() and it |
| 1680 | * may rebalance the tree from under us. So |
| 1681 | * restart the search from scratch. Returning |
| 1682 | * NULL would be safe too, but we'd generate |
| 1683 | * false negative insertions just because some |
| 1684 | * stable_node was stale. |
| 1685 | */ |
| 1686 | goto again; |
| 1687 | } |
| 1688 | |
| 1689 | ret = memcmp_pages(page, tree_page); |
| 1690 | put_page(tree_page); |
| 1691 | |
| 1692 | parent = *new; |
| 1693 | if (ret < 0) |
| 1694 | new = &parent->rb_left; |
| 1695 | else if (ret > 0) |
| 1696 | new = &parent->rb_right; |
| 1697 | else { |
| 1698 | if (page_node) { |
| 1699 | VM_BUG_ON(page_node->head != &migrate_nodes); |
| 1700 | /* |
| 1701 | * Test if the migrated page should be merged |
| 1702 | * into a stable node dup. If the mapcount is |
| 1703 | * 1 we can migrate it with another KSM page |
| 1704 | * without adding it to the chain. |
| 1705 | */ |
| 1706 | if (page_mapcount(page) > 1) |
| 1707 | goto chain_append; |
| 1708 | } |
| 1709 | |
| 1710 | if (!stable_node_dup) { |
| 1711 | /* |
| 1712 | * If the stable_node is a chain and |
| 1713 | * we got a payload match in memcmp |
| 1714 | * but we cannot merge the scanned |
| 1715 | * page in any of the existing |
| 1716 | * stable_node dups because they're |
| 1717 | * all full, we need to wait the |
| 1718 | * scanned page to find itself a match |
| 1719 | * in the unstable tree to create a |
| 1720 | * brand new KSM page to add later to |
| 1721 | * the dups of this stable_node. |
| 1722 | */ |
| 1723 | return NULL; |
| 1724 | } |
| 1725 | |
| 1726 | /* |
| 1727 | * Lock and unlock the stable_node's page (which |
| 1728 | * might already have been migrated) so that page |
| 1729 | * migration is sure to notice its raised count. |
| 1730 | * It would be more elegant to return stable_node |
| 1731 | * than kpage, but that involves more changes. |
| 1732 | */ |
| 1733 | tree_page = get_ksm_page(stable_node_dup, |
| 1734 | GET_KSM_PAGE_TRYLOCK); |
| 1735 | |
| 1736 | if (PTR_ERR(tree_page) == -EBUSY) |
| 1737 | return ERR_PTR(-EBUSY); |
| 1738 | |
| 1739 | if (unlikely(!tree_page)) |
| 1740 | /* |
| 1741 | * The tree may have been rebalanced, |
| 1742 | * so re-evaluate parent and new. |
| 1743 | */ |
| 1744 | goto again; |
| 1745 | unlock_page(tree_page); |
| 1746 | |
| 1747 | if (get_kpfn_nid(stable_node_dup->kpfn) != |
| 1748 | NUMA(stable_node_dup->nid)) { |
| 1749 | put_page(tree_page); |
| 1750 | goto replace; |
| 1751 | } |
| 1752 | return tree_page; |
| 1753 | } |
| 1754 | } |
| 1755 | |
| 1756 | if (!page_node) |
| 1757 | return NULL; |
| 1758 | |
| 1759 | list_del(&page_node->list); |
| 1760 | DO_NUMA(page_node->nid = nid); |
| 1761 | rb_link_node(&page_node->node, parent, new); |
| 1762 | rb_insert_color(&page_node->node, root); |
| 1763 | out: |
| 1764 | if (is_page_sharing_candidate(page_node)) { |
| 1765 | get_page(page); |
| 1766 | return page; |
| 1767 | } else |
| 1768 | return NULL; |
| 1769 | |
| 1770 | replace: |
| 1771 | /* |
| 1772 | * If stable_node was a chain and chain_prune collapsed it, |
| 1773 | * stable_node has been updated to be the new regular |
| 1774 | * stable_node. A collapse of the chain is indistinguishable |
| 1775 | * from the case there was no chain in the stable |
| 1776 | * rbtree. Otherwise stable_node is the chain and |
| 1777 | * stable_node_dup is the dup to replace. |
| 1778 | */ |
| 1779 | if (stable_node_dup == stable_node) { |
| 1780 | VM_BUG_ON(is_stable_node_chain(stable_node_dup)); |
| 1781 | VM_BUG_ON(is_stable_node_dup(stable_node_dup)); |
| 1782 | /* there is no chain */ |
| 1783 | if (page_node) { |
| 1784 | VM_BUG_ON(page_node->head != &migrate_nodes); |
| 1785 | list_del(&page_node->list); |
| 1786 | DO_NUMA(page_node->nid = nid); |
| 1787 | rb_replace_node(&stable_node_dup->node, |
| 1788 | &page_node->node, |
| 1789 | root); |
| 1790 | if (is_page_sharing_candidate(page_node)) |
| 1791 | get_page(page); |
| 1792 | else |
| 1793 | page = NULL; |
| 1794 | } else { |
| 1795 | rb_erase(&stable_node_dup->node, root); |
| 1796 | page = NULL; |
| 1797 | } |
| 1798 | } else { |
| 1799 | VM_BUG_ON(!is_stable_node_chain(stable_node)); |
| 1800 | __stable_node_dup_del(stable_node_dup); |
| 1801 | if (page_node) { |
| 1802 | VM_BUG_ON(page_node->head != &migrate_nodes); |
| 1803 | list_del(&page_node->list); |
| 1804 | DO_NUMA(page_node->nid = nid); |
| 1805 | stable_node_chain_add_dup(page_node, stable_node); |
| 1806 | if (is_page_sharing_candidate(page_node)) |
| 1807 | get_page(page); |
| 1808 | else |
| 1809 | page = NULL; |
| 1810 | } else { |
| 1811 | page = NULL; |
| 1812 | } |
| 1813 | } |
| 1814 | stable_node_dup->head = &migrate_nodes; |
| 1815 | list_add(&stable_node_dup->list, stable_node_dup->head); |
| 1816 | return page; |
| 1817 | |
| 1818 | chain_append: |
| 1819 | /* stable_node_dup could be null if it reached the limit */ |
| 1820 | if (!stable_node_dup) |
| 1821 | stable_node_dup = stable_node_any; |
| 1822 | /* |
| 1823 | * If stable_node was a chain and chain_prune collapsed it, |
| 1824 | * stable_node has been updated to be the new regular |
| 1825 | * stable_node. A collapse of the chain is indistinguishable |
| 1826 | * from the case there was no chain in the stable |
| 1827 | * rbtree. Otherwise stable_node is the chain and |
| 1828 | * stable_node_dup is the dup to replace. |
| 1829 | */ |
| 1830 | if (stable_node_dup == stable_node) { |
| 1831 | VM_BUG_ON(is_stable_node_dup(stable_node_dup)); |
| 1832 | /* chain is missing so create it */ |
| 1833 | stable_node = alloc_stable_node_chain(stable_node_dup, |
| 1834 | root); |
| 1835 | if (!stable_node) |
| 1836 | return NULL; |
| 1837 | } |
| 1838 | /* |
| 1839 | * Add this stable_node dup that was |
| 1840 | * migrated to the stable_node chain |
| 1841 | * of the current nid for this page |
| 1842 | * content. |
| 1843 | */ |
| 1844 | VM_BUG_ON(!is_stable_node_dup(stable_node_dup)); |
| 1845 | VM_BUG_ON(page_node->head != &migrate_nodes); |
| 1846 | list_del(&page_node->list); |
| 1847 | DO_NUMA(page_node->nid = nid); |
| 1848 | stable_node_chain_add_dup(page_node, stable_node); |
| 1849 | goto out; |
| 1850 | } |
| 1851 | |
| 1852 | /* |
| 1853 | * stable_tree_insert - insert stable tree node pointing to new ksm page |
| 1854 | * into the stable tree. |
| 1855 | * |
| 1856 | * This function returns the stable tree node just allocated on success, |
| 1857 | * NULL otherwise. |
| 1858 | */ |
| 1859 | static struct ksm_stable_node *stable_tree_insert(struct page *kpage) |
| 1860 | { |
| 1861 | int nid; |
| 1862 | unsigned long kpfn; |
| 1863 | struct rb_root *root; |
| 1864 | struct rb_node **new; |
| 1865 | struct rb_node *parent; |
| 1866 | struct ksm_stable_node *stable_node, *stable_node_dup, *stable_node_any; |
| 1867 | bool need_chain = false; |
| 1868 | |
| 1869 | kpfn = page_to_pfn(kpage); |
| 1870 | nid = get_kpfn_nid(kpfn); |
| 1871 | root = root_stable_tree + nid; |
| 1872 | again: |
| 1873 | parent = NULL; |
| 1874 | new = &root->rb_node; |
| 1875 | |
| 1876 | while (*new) { |
| 1877 | struct page *tree_page; |
| 1878 | int ret; |
| 1879 | |
| 1880 | cond_resched(); |
| 1881 | stable_node = rb_entry(*new, struct ksm_stable_node, node); |
| 1882 | stable_node_any = NULL; |
| 1883 | tree_page = chain(&stable_node_dup, stable_node, root); |
| 1884 | if (!stable_node_dup) { |
| 1885 | /* |
| 1886 | * Either all stable_node dups were full in |
| 1887 | * this stable_node chain, or this chain was |
| 1888 | * empty and should be rb_erased. |
| 1889 | */ |
| 1890 | stable_node_any = stable_node_dup_any(stable_node, |
| 1891 | root); |
| 1892 | if (!stable_node_any) { |
| 1893 | /* rb_erase just run */ |
| 1894 | goto again; |
| 1895 | } |
| 1896 | /* |
| 1897 | * Take any of the stable_node dups page of |
| 1898 | * this stable_node chain to let the tree walk |
| 1899 | * continue. All KSM pages belonging to the |
| 1900 | * stable_node dups in a stable_node chain |
| 1901 | * have the same content and they're |
| 1902 | * write protected at all times. Any will work |
| 1903 | * fine to continue the walk. |
| 1904 | */ |
| 1905 | tree_page = get_ksm_page(stable_node_any, |
| 1906 | GET_KSM_PAGE_NOLOCK); |
| 1907 | } |
| 1908 | VM_BUG_ON(!stable_node_dup ^ !!stable_node_any); |
| 1909 | if (!tree_page) { |
| 1910 | /* |
| 1911 | * If we walked over a stale stable_node, |
| 1912 | * get_ksm_page() will call rb_erase() and it |
| 1913 | * may rebalance the tree from under us. So |
| 1914 | * restart the search from scratch. Returning |
| 1915 | * NULL would be safe too, but we'd generate |
| 1916 | * false negative insertions just because some |
| 1917 | * stable_node was stale. |
| 1918 | */ |
| 1919 | goto again; |
| 1920 | } |
| 1921 | |
| 1922 | ret = memcmp_pages(kpage, tree_page); |
| 1923 | put_page(tree_page); |
| 1924 | |
| 1925 | parent = *new; |
| 1926 | if (ret < 0) |
| 1927 | new = &parent->rb_left; |
| 1928 | else if (ret > 0) |
| 1929 | new = &parent->rb_right; |
| 1930 | else { |
| 1931 | need_chain = true; |
| 1932 | break; |
| 1933 | } |
| 1934 | } |
| 1935 | |
| 1936 | stable_node_dup = alloc_stable_node(); |
| 1937 | if (!stable_node_dup) |
| 1938 | return NULL; |
| 1939 | |
| 1940 | INIT_HLIST_HEAD(&stable_node_dup->hlist); |
| 1941 | stable_node_dup->kpfn = kpfn; |
| 1942 | set_page_stable_node(kpage, stable_node_dup); |
| 1943 | stable_node_dup->rmap_hlist_len = 0; |
| 1944 | DO_NUMA(stable_node_dup->nid = nid); |
| 1945 | if (!need_chain) { |
| 1946 | rb_link_node(&stable_node_dup->node, parent, new); |
| 1947 | rb_insert_color(&stable_node_dup->node, root); |
| 1948 | } else { |
| 1949 | if (!is_stable_node_chain(stable_node)) { |
| 1950 | struct ksm_stable_node *orig = stable_node; |
| 1951 | /* chain is missing so create it */ |
| 1952 | stable_node = alloc_stable_node_chain(orig, root); |
| 1953 | if (!stable_node) { |
| 1954 | free_stable_node(stable_node_dup); |
| 1955 | return NULL; |
| 1956 | } |
| 1957 | } |
| 1958 | stable_node_chain_add_dup(stable_node_dup, stable_node); |
| 1959 | } |
| 1960 | |
| 1961 | return stable_node_dup; |
| 1962 | } |
| 1963 | |
| 1964 | /* |
| 1965 | * unstable_tree_search_insert - search for identical page, |
| 1966 | * else insert rmap_item into the unstable tree. |
| 1967 | * |
| 1968 | * This function searches for a page in the unstable tree identical to the |
| 1969 | * page currently being scanned; and if no identical page is found in the |
| 1970 | * tree, we insert rmap_item as a new object into the unstable tree. |
| 1971 | * |
| 1972 | * This function returns pointer to rmap_item found to be identical |
| 1973 | * to the currently scanned page, NULL otherwise. |
| 1974 | * |
| 1975 | * This function does both searching and inserting, because they share |
| 1976 | * the same walking algorithm in an rbtree. |
| 1977 | */ |
| 1978 | static |
| 1979 | struct ksm_rmap_item *unstable_tree_search_insert(struct ksm_rmap_item *rmap_item, |
| 1980 | struct page *page, |
| 1981 | struct page **tree_pagep) |
| 1982 | { |
| 1983 | struct rb_node **new; |
| 1984 | struct rb_root *root; |
| 1985 | struct rb_node *parent = NULL; |
| 1986 | int nid; |
| 1987 | |
| 1988 | nid = get_kpfn_nid(page_to_pfn(page)); |
| 1989 | root = root_unstable_tree + nid; |
| 1990 | new = &root->rb_node; |
| 1991 | |
| 1992 | while (*new) { |
| 1993 | struct ksm_rmap_item *tree_rmap_item; |
| 1994 | struct page *tree_page; |
| 1995 | int ret; |
| 1996 | |
| 1997 | cond_resched(); |
| 1998 | tree_rmap_item = rb_entry(*new, struct ksm_rmap_item, node); |
| 1999 | tree_page = get_mergeable_page(tree_rmap_item); |
| 2000 | if (!tree_page) |
| 2001 | return NULL; |
| 2002 | |
| 2003 | /* |
| 2004 | * Don't substitute a ksm page for a forked page. |
| 2005 | */ |
| 2006 | if (page == tree_page) { |
| 2007 | put_page(tree_page); |
| 2008 | return NULL; |
| 2009 | } |
| 2010 | |
| 2011 | ret = memcmp_pages(page, tree_page); |
| 2012 | |
| 2013 | parent = *new; |
| 2014 | if (ret < 0) { |
| 2015 | put_page(tree_page); |
| 2016 | new = &parent->rb_left; |
| 2017 | } else if (ret > 0) { |
| 2018 | put_page(tree_page); |
| 2019 | new = &parent->rb_right; |
| 2020 | } else if (!ksm_merge_across_nodes && |
| 2021 | page_to_nid(tree_page) != nid) { |
| 2022 | /* |
| 2023 | * If tree_page has been migrated to another NUMA node, |
| 2024 | * it will be flushed out and put in the right unstable |
| 2025 | * tree next time: only merge with it when across_nodes. |
| 2026 | */ |
| 2027 | put_page(tree_page); |
| 2028 | return NULL; |
| 2029 | } else { |
| 2030 | *tree_pagep = tree_page; |
| 2031 | return tree_rmap_item; |
| 2032 | } |
| 2033 | } |
| 2034 | |
| 2035 | rmap_item->address |= UNSTABLE_FLAG; |
| 2036 | rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK); |
| 2037 | DO_NUMA(rmap_item->nid = nid); |
| 2038 | rb_link_node(&rmap_item->node, parent, new); |
| 2039 | rb_insert_color(&rmap_item->node, root); |
| 2040 | |
| 2041 | ksm_pages_unshared++; |
| 2042 | return NULL; |
| 2043 | } |
| 2044 | |
| 2045 | /* |
| 2046 | * stable_tree_append - add another rmap_item to the linked list of |
| 2047 | * rmap_items hanging off a given node of the stable tree, all sharing |
| 2048 | * the same ksm page. |
| 2049 | */ |
| 2050 | static void stable_tree_append(struct ksm_rmap_item *rmap_item, |
| 2051 | struct ksm_stable_node *stable_node, |
| 2052 | bool max_page_sharing_bypass) |
| 2053 | { |
| 2054 | /* |
| 2055 | * rmap won't find this mapping if we don't insert the |
| 2056 | * rmap_item in the right stable_node |
| 2057 | * duplicate. page_migration could break later if rmap breaks, |
| 2058 | * so we can as well crash here. We really need to check for |
| 2059 | * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check |
| 2060 | * for other negative values as an underflow if detected here |
| 2061 | * for the first time (and not when decreasing rmap_hlist_len) |
| 2062 | * would be sign of memory corruption in the stable_node. |
| 2063 | */ |
| 2064 | BUG_ON(stable_node->rmap_hlist_len < 0); |
| 2065 | |
| 2066 | stable_node->rmap_hlist_len++; |
| 2067 | if (!max_page_sharing_bypass) |
| 2068 | /* possibly non fatal but unexpected overflow, only warn */ |
| 2069 | WARN_ON_ONCE(stable_node->rmap_hlist_len > |
| 2070 | ksm_max_page_sharing); |
| 2071 | |
| 2072 | rmap_item->head = stable_node; |
| 2073 | rmap_item->address |= STABLE_FLAG; |
| 2074 | hlist_add_head(&rmap_item->hlist, &stable_node->hlist); |
| 2075 | |
| 2076 | if (rmap_item->hlist.next) |
| 2077 | ksm_pages_sharing++; |
| 2078 | else |
| 2079 | ksm_pages_shared++; |
| 2080 | |
| 2081 | rmap_item->mm->ksm_merging_pages++; |
| 2082 | } |
| 2083 | |
| 2084 | /* |
| 2085 | * cmp_and_merge_page - first see if page can be merged into the stable tree; |
| 2086 | * if not, compare checksum to previous and if it's the same, see if page can |
| 2087 | * be inserted into the unstable tree, or merged with a page already there and |
| 2088 | * both transferred to the stable tree. |
| 2089 | * |
| 2090 | * @page: the page that we are searching identical page to. |
| 2091 | * @rmap_item: the reverse mapping into the virtual address of this page |
| 2092 | */ |
| 2093 | static void cmp_and_merge_page(struct page *page, struct ksm_rmap_item *rmap_item) |
| 2094 | { |
| 2095 | struct mm_struct *mm = rmap_item->mm; |
| 2096 | struct ksm_rmap_item *tree_rmap_item; |
| 2097 | struct page *tree_page = NULL; |
| 2098 | struct ksm_stable_node *stable_node; |
| 2099 | struct page *kpage; |
| 2100 | unsigned int checksum; |
| 2101 | int err; |
| 2102 | bool max_page_sharing_bypass = false; |
| 2103 | |
| 2104 | stable_node = page_stable_node(page); |
| 2105 | if (stable_node) { |
| 2106 | if (stable_node->head != &migrate_nodes && |
| 2107 | get_kpfn_nid(READ_ONCE(stable_node->kpfn)) != |
| 2108 | NUMA(stable_node->nid)) { |
| 2109 | stable_node_dup_del(stable_node); |
| 2110 | stable_node->head = &migrate_nodes; |
| 2111 | list_add(&stable_node->list, stable_node->head); |
| 2112 | } |
| 2113 | if (stable_node->head != &migrate_nodes && |
| 2114 | rmap_item->head == stable_node) |
| 2115 | return; |
| 2116 | /* |
| 2117 | * If it's a KSM fork, allow it to go over the sharing limit |
| 2118 | * without warnings. |
| 2119 | */ |
| 2120 | if (!is_page_sharing_candidate(stable_node)) |
| 2121 | max_page_sharing_bypass = true; |
| 2122 | } |
| 2123 | |
| 2124 | /* We first start with searching the page inside the stable tree */ |
| 2125 | kpage = stable_tree_search(page); |
| 2126 | if (kpage == page && rmap_item->head == stable_node) { |
| 2127 | put_page(kpage); |
| 2128 | return; |
| 2129 | } |
| 2130 | |
| 2131 | remove_rmap_item_from_tree(rmap_item); |
| 2132 | |
| 2133 | if (kpage) { |
| 2134 | if (PTR_ERR(kpage) == -EBUSY) |
| 2135 | return; |
| 2136 | |
| 2137 | err = try_to_merge_with_ksm_page(rmap_item, page, kpage); |
| 2138 | if (!err) { |
| 2139 | /* |
| 2140 | * The page was successfully merged: |
| 2141 | * add its rmap_item to the stable tree. |
| 2142 | */ |
| 2143 | lock_page(kpage); |
| 2144 | stable_tree_append(rmap_item, page_stable_node(kpage), |
| 2145 | max_page_sharing_bypass); |
| 2146 | unlock_page(kpage); |
| 2147 | } |
| 2148 | put_page(kpage); |
| 2149 | return; |
| 2150 | } |
| 2151 | |
| 2152 | /* |
| 2153 | * If the hash value of the page has changed from the last time |
| 2154 | * we calculated it, this page is changing frequently: therefore we |
| 2155 | * don't want to insert it in the unstable tree, and we don't want |
| 2156 | * to waste our time searching for something identical to it there. |
| 2157 | */ |
| 2158 | checksum = calc_checksum(page); |
| 2159 | if (rmap_item->oldchecksum != checksum) { |
| 2160 | rmap_item->oldchecksum = checksum; |
| 2161 | return; |
| 2162 | } |
| 2163 | |
| 2164 | /* |
| 2165 | * Same checksum as an empty page. We attempt to merge it with the |
| 2166 | * appropriate zero page if the user enabled this via sysfs. |
| 2167 | */ |
| 2168 | if (ksm_use_zero_pages && (checksum == zero_checksum)) { |
| 2169 | struct vm_area_struct *vma; |
| 2170 | |
| 2171 | mmap_read_lock(mm); |
| 2172 | vma = find_mergeable_vma(mm, rmap_item->address); |
| 2173 | if (vma) { |
| 2174 | err = try_to_merge_one_page(vma, page, |
| 2175 | ZERO_PAGE(rmap_item->address)); |
| 2176 | trace_ksm_merge_one_page( |
| 2177 | page_to_pfn(ZERO_PAGE(rmap_item->address)), |
| 2178 | rmap_item, mm, err); |
| 2179 | } else { |
| 2180 | /* |
| 2181 | * If the vma is out of date, we do not need to |
| 2182 | * continue. |
| 2183 | */ |
| 2184 | err = 0; |
| 2185 | } |
| 2186 | mmap_read_unlock(mm); |
| 2187 | /* |
| 2188 | * In case of failure, the page was not really empty, so we |
| 2189 | * need to continue. Otherwise we're done. |
| 2190 | */ |
| 2191 | if (!err) |
| 2192 | return; |
| 2193 | } |
| 2194 | tree_rmap_item = |
| 2195 | unstable_tree_search_insert(rmap_item, page, &tree_page); |
| 2196 | if (tree_rmap_item) { |
| 2197 | bool split; |
| 2198 | |
| 2199 | kpage = try_to_merge_two_pages(rmap_item, page, |
| 2200 | tree_rmap_item, tree_page); |
| 2201 | /* |
| 2202 | * If both pages we tried to merge belong to the same compound |
| 2203 | * page, then we actually ended up increasing the reference |
| 2204 | * count of the same compound page twice, and split_huge_page |
| 2205 | * failed. |
| 2206 | * Here we set a flag if that happened, and we use it later to |
| 2207 | * try split_huge_page again. Since we call put_page right |
| 2208 | * afterwards, the reference count will be correct and |
| 2209 | * split_huge_page should succeed. |
| 2210 | */ |
| 2211 | split = PageTransCompound(page) |
| 2212 | && compound_head(page) == compound_head(tree_page); |
| 2213 | put_page(tree_page); |
| 2214 | if (kpage) { |
| 2215 | /* |
| 2216 | * The pages were successfully merged: insert new |
| 2217 | * node in the stable tree and add both rmap_items. |
| 2218 | */ |
| 2219 | lock_page(kpage); |
| 2220 | stable_node = stable_tree_insert(kpage); |
| 2221 | if (stable_node) { |
| 2222 | stable_tree_append(tree_rmap_item, stable_node, |
| 2223 | false); |
| 2224 | stable_tree_append(rmap_item, stable_node, |
| 2225 | false); |
| 2226 | } |
| 2227 | unlock_page(kpage); |
| 2228 | |
| 2229 | /* |
| 2230 | * If we fail to insert the page into the stable tree, |
| 2231 | * we will have 2 virtual addresses that are pointing |
| 2232 | * to a ksm page left outside the stable tree, |
| 2233 | * in which case we need to break_cow on both. |
| 2234 | */ |
| 2235 | if (!stable_node) { |
| 2236 | break_cow(tree_rmap_item); |
| 2237 | break_cow(rmap_item); |
| 2238 | } |
| 2239 | } else if (split) { |
| 2240 | /* |
| 2241 | * We are here if we tried to merge two pages and |
| 2242 | * failed because they both belonged to the same |
| 2243 | * compound page. We will split the page now, but no |
| 2244 | * merging will take place. |
| 2245 | * We do not want to add the cost of a full lock; if |
| 2246 | * the page is locked, it is better to skip it and |
| 2247 | * perhaps try again later. |
| 2248 | */ |
| 2249 | if (!trylock_page(page)) |
| 2250 | return; |
| 2251 | split_huge_page(page); |
| 2252 | unlock_page(page); |
| 2253 | } |
| 2254 | } |
| 2255 | } |
| 2256 | |
| 2257 | static struct ksm_rmap_item *get_next_rmap_item(struct ksm_mm_slot *mm_slot, |
| 2258 | struct ksm_rmap_item **rmap_list, |
| 2259 | unsigned long addr) |
| 2260 | { |
| 2261 | struct ksm_rmap_item *rmap_item; |
| 2262 | |
| 2263 | while (*rmap_list) { |
| 2264 | rmap_item = *rmap_list; |
| 2265 | if ((rmap_item->address & PAGE_MASK) == addr) |
| 2266 | return rmap_item; |
| 2267 | if (rmap_item->address > addr) |
| 2268 | break; |
| 2269 | *rmap_list = rmap_item->rmap_list; |
| 2270 | remove_rmap_item_from_tree(rmap_item); |
| 2271 | free_rmap_item(rmap_item); |
| 2272 | } |
| 2273 | |
| 2274 | rmap_item = alloc_rmap_item(); |
| 2275 | if (rmap_item) { |
| 2276 | /* It has already been zeroed */ |
| 2277 | rmap_item->mm = mm_slot->slot.mm; |
| 2278 | rmap_item->mm->ksm_rmap_items++; |
| 2279 | rmap_item->address = addr; |
| 2280 | rmap_item->rmap_list = *rmap_list; |
| 2281 | *rmap_list = rmap_item; |
| 2282 | } |
| 2283 | return rmap_item; |
| 2284 | } |
| 2285 | |
| 2286 | static struct ksm_rmap_item *scan_get_next_rmap_item(struct page **page) |
| 2287 | { |
| 2288 | struct mm_struct *mm; |
| 2289 | struct ksm_mm_slot *mm_slot; |
| 2290 | struct mm_slot *slot; |
| 2291 | struct vm_area_struct *vma; |
| 2292 | struct ksm_rmap_item *rmap_item; |
| 2293 | struct vma_iterator vmi; |
| 2294 | int nid; |
| 2295 | |
| 2296 | if (list_empty(&ksm_mm_head.slot.mm_node)) |
| 2297 | return NULL; |
| 2298 | |
| 2299 | mm_slot = ksm_scan.mm_slot; |
| 2300 | if (mm_slot == &ksm_mm_head) { |
| 2301 | trace_ksm_start_scan(ksm_scan.seqnr, ksm_rmap_items); |
| 2302 | |
| 2303 | /* |
| 2304 | * A number of pages can hang around indefinitely on per-cpu |
| 2305 | * pagevecs, raised page count preventing write_protect_page |
| 2306 | * from merging them. Though it doesn't really matter much, |
| 2307 | * it is puzzling to see some stuck in pages_volatile until |
| 2308 | * other activity jostles them out, and they also prevented |
| 2309 | * LTP's KSM test from succeeding deterministically; so drain |
| 2310 | * them here (here rather than on entry to ksm_do_scan(), |
| 2311 | * so we don't IPI too often when pages_to_scan is set low). |
| 2312 | */ |
| 2313 | lru_add_drain_all(); |
| 2314 | |
| 2315 | /* |
| 2316 | * Whereas stale stable_nodes on the stable_tree itself |
| 2317 | * get pruned in the regular course of stable_tree_search(), |
| 2318 | * those moved out to the migrate_nodes list can accumulate: |
| 2319 | * so prune them once before each full scan. |
| 2320 | */ |
| 2321 | if (!ksm_merge_across_nodes) { |
| 2322 | struct ksm_stable_node *stable_node, *next; |
| 2323 | struct page *page; |
| 2324 | |
| 2325 | list_for_each_entry_safe(stable_node, next, |
| 2326 | &migrate_nodes, list) { |
| 2327 | page = get_ksm_page(stable_node, |
| 2328 | GET_KSM_PAGE_NOLOCK); |
| 2329 | if (page) |
| 2330 | put_page(page); |
| 2331 | cond_resched(); |
| 2332 | } |
| 2333 | } |
| 2334 | |
| 2335 | for (nid = 0; nid < ksm_nr_node_ids; nid++) |
| 2336 | root_unstable_tree[nid] = RB_ROOT; |
| 2337 | |
| 2338 | spin_lock(&ksm_mmlist_lock); |
| 2339 | slot = list_entry(mm_slot->slot.mm_node.next, |
| 2340 | struct mm_slot, mm_node); |
| 2341 | mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); |
| 2342 | ksm_scan.mm_slot = mm_slot; |
| 2343 | spin_unlock(&ksm_mmlist_lock); |
| 2344 | /* |
| 2345 | * Although we tested list_empty() above, a racing __ksm_exit |
| 2346 | * of the last mm on the list may have removed it since then. |
| 2347 | */ |
| 2348 | if (mm_slot == &ksm_mm_head) |
| 2349 | return NULL; |
| 2350 | next_mm: |
| 2351 | ksm_scan.address = 0; |
| 2352 | ksm_scan.rmap_list = &mm_slot->rmap_list; |
| 2353 | } |
| 2354 | |
| 2355 | slot = &mm_slot->slot; |
| 2356 | mm = slot->mm; |
| 2357 | vma_iter_init(&vmi, mm, ksm_scan.address); |
| 2358 | |
| 2359 | mmap_read_lock(mm); |
| 2360 | if (ksm_test_exit(mm)) |
| 2361 | goto no_vmas; |
| 2362 | |
| 2363 | for_each_vma(vmi, vma) { |
| 2364 | if (!(vma->vm_flags & VM_MERGEABLE)) |
| 2365 | continue; |
| 2366 | if (ksm_scan.address < vma->vm_start) |
| 2367 | ksm_scan.address = vma->vm_start; |
| 2368 | if (!vma->anon_vma) |
| 2369 | ksm_scan.address = vma->vm_end; |
| 2370 | |
| 2371 | while (ksm_scan.address < vma->vm_end) { |
| 2372 | if (ksm_test_exit(mm)) |
| 2373 | break; |
| 2374 | *page = follow_page(vma, ksm_scan.address, FOLL_GET); |
| 2375 | if (IS_ERR_OR_NULL(*page)) { |
| 2376 | ksm_scan.address += PAGE_SIZE; |
| 2377 | cond_resched(); |
| 2378 | continue; |
| 2379 | } |
| 2380 | if (is_zone_device_page(*page)) |
| 2381 | goto next_page; |
| 2382 | if (PageAnon(*page)) { |
| 2383 | flush_anon_page(vma, *page, ksm_scan.address); |
| 2384 | flush_dcache_page(*page); |
| 2385 | rmap_item = get_next_rmap_item(mm_slot, |
| 2386 | ksm_scan.rmap_list, ksm_scan.address); |
| 2387 | if (rmap_item) { |
| 2388 | ksm_scan.rmap_list = |
| 2389 | &rmap_item->rmap_list; |
| 2390 | ksm_scan.address += PAGE_SIZE; |
| 2391 | } else |
| 2392 | put_page(*page); |
| 2393 | mmap_read_unlock(mm); |
| 2394 | return rmap_item; |
| 2395 | } |
| 2396 | next_page: |
| 2397 | put_page(*page); |
| 2398 | ksm_scan.address += PAGE_SIZE; |
| 2399 | cond_resched(); |
| 2400 | } |
| 2401 | } |
| 2402 | |
| 2403 | if (ksm_test_exit(mm)) { |
| 2404 | no_vmas: |
| 2405 | ksm_scan.address = 0; |
| 2406 | ksm_scan.rmap_list = &mm_slot->rmap_list; |
| 2407 | } |
| 2408 | /* |
| 2409 | * Nuke all the rmap_items that are above this current rmap: |
| 2410 | * because there were no VM_MERGEABLE vmas with such addresses. |
| 2411 | */ |
| 2412 | remove_trailing_rmap_items(ksm_scan.rmap_list); |
| 2413 | |
| 2414 | spin_lock(&ksm_mmlist_lock); |
| 2415 | slot = list_entry(mm_slot->slot.mm_node.next, |
| 2416 | struct mm_slot, mm_node); |
| 2417 | ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); |
| 2418 | if (ksm_scan.address == 0) { |
| 2419 | /* |
| 2420 | * We've completed a full scan of all vmas, holding mmap_lock |
| 2421 | * throughout, and found no VM_MERGEABLE: so do the same as |
| 2422 | * __ksm_exit does to remove this mm from all our lists now. |
| 2423 | * This applies either when cleaning up after __ksm_exit |
| 2424 | * (but beware: we can reach here even before __ksm_exit), |
| 2425 | * or when all VM_MERGEABLE areas have been unmapped (and |
| 2426 | * mmap_lock then protects against race with MADV_MERGEABLE). |
| 2427 | */ |
| 2428 | hash_del(&mm_slot->slot.hash); |
| 2429 | list_del(&mm_slot->slot.mm_node); |
| 2430 | spin_unlock(&ksm_mmlist_lock); |
| 2431 | |
| 2432 | mm_slot_free(mm_slot_cache, mm_slot); |
| 2433 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); |
| 2434 | clear_bit(MMF_VM_MERGE_ANY, &mm->flags); |
| 2435 | mmap_read_unlock(mm); |
| 2436 | mmdrop(mm); |
| 2437 | } else { |
| 2438 | mmap_read_unlock(mm); |
| 2439 | /* |
| 2440 | * mmap_read_unlock(mm) first because after |
| 2441 | * spin_unlock(&ksm_mmlist_lock) run, the "mm" may |
| 2442 | * already have been freed under us by __ksm_exit() |
| 2443 | * because the "mm_slot" is still hashed and |
| 2444 | * ksm_scan.mm_slot doesn't point to it anymore. |
| 2445 | */ |
| 2446 | spin_unlock(&ksm_mmlist_lock); |
| 2447 | } |
| 2448 | |
| 2449 | /* Repeat until we've completed scanning the whole list */ |
| 2450 | mm_slot = ksm_scan.mm_slot; |
| 2451 | if (mm_slot != &ksm_mm_head) |
| 2452 | goto next_mm; |
| 2453 | |
| 2454 | trace_ksm_stop_scan(ksm_scan.seqnr, ksm_rmap_items); |
| 2455 | ksm_scan.seqnr++; |
| 2456 | return NULL; |
| 2457 | } |
| 2458 | |
| 2459 | /** |
| 2460 | * ksm_do_scan - the ksm scanner main worker function. |
| 2461 | * @scan_npages: number of pages we want to scan before we return. |
| 2462 | */ |
| 2463 | static void ksm_do_scan(unsigned int scan_npages) |
| 2464 | { |
| 2465 | struct ksm_rmap_item *rmap_item; |
| 2466 | struct page *page; |
| 2467 | |
| 2468 | while (scan_npages-- && likely(!freezing(current))) { |
| 2469 | cond_resched(); |
| 2470 | rmap_item = scan_get_next_rmap_item(&page); |
| 2471 | if (!rmap_item) |
| 2472 | return; |
| 2473 | cmp_and_merge_page(page, rmap_item); |
| 2474 | put_page(page); |
| 2475 | } |
| 2476 | } |
| 2477 | |
| 2478 | static int ksmd_should_run(void) |
| 2479 | { |
| 2480 | return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.slot.mm_node); |
| 2481 | } |
| 2482 | |
| 2483 | static int ksm_scan_thread(void *nothing) |
| 2484 | { |
| 2485 | unsigned int sleep_ms; |
| 2486 | |
| 2487 | set_freezable(); |
| 2488 | set_user_nice(current, 5); |
| 2489 | |
| 2490 | while (!kthread_should_stop()) { |
| 2491 | mutex_lock(&ksm_thread_mutex); |
| 2492 | wait_while_offlining(); |
| 2493 | if (ksmd_should_run()) |
| 2494 | ksm_do_scan(ksm_thread_pages_to_scan); |
| 2495 | mutex_unlock(&ksm_thread_mutex); |
| 2496 | |
| 2497 | try_to_freeze(); |
| 2498 | |
| 2499 | if (ksmd_should_run()) { |
| 2500 | sleep_ms = READ_ONCE(ksm_thread_sleep_millisecs); |
| 2501 | wait_event_interruptible_timeout(ksm_iter_wait, |
| 2502 | sleep_ms != READ_ONCE(ksm_thread_sleep_millisecs), |
| 2503 | msecs_to_jiffies(sleep_ms)); |
| 2504 | } else { |
| 2505 | wait_event_freezable(ksm_thread_wait, |
| 2506 | ksmd_should_run() || kthread_should_stop()); |
| 2507 | } |
| 2508 | } |
| 2509 | return 0; |
| 2510 | } |
| 2511 | |
| 2512 | static void __ksm_add_vma(struct vm_area_struct *vma) |
| 2513 | { |
| 2514 | unsigned long vm_flags = vma->vm_flags; |
| 2515 | |
| 2516 | if (vm_flags & VM_MERGEABLE) |
| 2517 | return; |
| 2518 | |
| 2519 | if (vma_ksm_compatible(vma)) |
| 2520 | vm_flags_set(vma, VM_MERGEABLE); |
| 2521 | } |
| 2522 | |
| 2523 | static int __ksm_del_vma(struct vm_area_struct *vma) |
| 2524 | { |
| 2525 | int err; |
| 2526 | |
| 2527 | if (!(vma->vm_flags & VM_MERGEABLE)) |
| 2528 | return 0; |
| 2529 | |
| 2530 | if (vma->anon_vma) { |
| 2531 | err = unmerge_ksm_pages(vma, vma->vm_start, vma->vm_end); |
| 2532 | if (err) |
| 2533 | return err; |
| 2534 | } |
| 2535 | |
| 2536 | vm_flags_clear(vma, VM_MERGEABLE); |
| 2537 | return 0; |
| 2538 | } |
| 2539 | /** |
| 2540 | * ksm_add_vma - Mark vma as mergeable if compatible |
| 2541 | * |
| 2542 | * @vma: Pointer to vma |
| 2543 | */ |
| 2544 | void ksm_add_vma(struct vm_area_struct *vma) |
| 2545 | { |
| 2546 | struct mm_struct *mm = vma->vm_mm; |
| 2547 | |
| 2548 | if (test_bit(MMF_VM_MERGE_ANY, &mm->flags)) |
| 2549 | __ksm_add_vma(vma); |
| 2550 | } |
| 2551 | |
| 2552 | static void ksm_add_vmas(struct mm_struct *mm) |
| 2553 | { |
| 2554 | struct vm_area_struct *vma; |
| 2555 | |
| 2556 | VMA_ITERATOR(vmi, mm, 0); |
| 2557 | for_each_vma(vmi, vma) |
| 2558 | __ksm_add_vma(vma); |
| 2559 | } |
| 2560 | |
| 2561 | static int ksm_del_vmas(struct mm_struct *mm) |
| 2562 | { |
| 2563 | struct vm_area_struct *vma; |
| 2564 | int err; |
| 2565 | |
| 2566 | VMA_ITERATOR(vmi, mm, 0); |
| 2567 | for_each_vma(vmi, vma) { |
| 2568 | err = __ksm_del_vma(vma); |
| 2569 | if (err) |
| 2570 | return err; |
| 2571 | } |
| 2572 | return 0; |
| 2573 | } |
| 2574 | |
| 2575 | /** |
| 2576 | * ksm_enable_merge_any - Add mm to mm ksm list and enable merging on all |
| 2577 | * compatible VMA's |
| 2578 | * |
| 2579 | * @mm: Pointer to mm |
| 2580 | * |
| 2581 | * Returns 0 on success, otherwise error code |
| 2582 | */ |
| 2583 | int ksm_enable_merge_any(struct mm_struct *mm) |
| 2584 | { |
| 2585 | int err; |
| 2586 | |
| 2587 | if (test_bit(MMF_VM_MERGE_ANY, &mm->flags)) |
| 2588 | return 0; |
| 2589 | |
| 2590 | if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) { |
| 2591 | err = __ksm_enter(mm); |
| 2592 | if (err) |
| 2593 | return err; |
| 2594 | } |
| 2595 | |
| 2596 | set_bit(MMF_VM_MERGE_ANY, &mm->flags); |
| 2597 | ksm_add_vmas(mm); |
| 2598 | |
| 2599 | return 0; |
| 2600 | } |
| 2601 | |
| 2602 | /** |
| 2603 | * ksm_disable_merge_any - Disable merging on all compatible VMA's of the mm, |
| 2604 | * previously enabled via ksm_enable_merge_any(). |
| 2605 | * |
| 2606 | * Disabling merging implies unmerging any merged pages, like setting |
| 2607 | * MADV_UNMERGEABLE would. If unmerging fails, the whole operation fails and |
| 2608 | * merging on all compatible VMA's remains enabled. |
| 2609 | * |
| 2610 | * @mm: Pointer to mm |
| 2611 | * |
| 2612 | * Returns 0 on success, otherwise error code |
| 2613 | */ |
| 2614 | int ksm_disable_merge_any(struct mm_struct *mm) |
| 2615 | { |
| 2616 | int err; |
| 2617 | |
| 2618 | if (!test_bit(MMF_VM_MERGE_ANY, &mm->flags)) |
| 2619 | return 0; |
| 2620 | |
| 2621 | err = ksm_del_vmas(mm); |
| 2622 | if (err) { |
| 2623 | ksm_add_vmas(mm); |
| 2624 | return err; |
| 2625 | } |
| 2626 | |
| 2627 | clear_bit(MMF_VM_MERGE_ANY, &mm->flags); |
| 2628 | return 0; |
| 2629 | } |
| 2630 | |
| 2631 | int ksm_disable(struct mm_struct *mm) |
| 2632 | { |
| 2633 | mmap_assert_write_locked(mm); |
| 2634 | |
| 2635 | if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) |
| 2636 | return 0; |
| 2637 | if (test_bit(MMF_VM_MERGE_ANY, &mm->flags)) |
| 2638 | return ksm_disable_merge_any(mm); |
| 2639 | return ksm_del_vmas(mm); |
| 2640 | } |
| 2641 | |
| 2642 | int ksm_madvise(struct vm_area_struct *vma, unsigned long start, |
| 2643 | unsigned long end, int advice, unsigned long *vm_flags) |
| 2644 | { |
| 2645 | struct mm_struct *mm = vma->vm_mm; |
| 2646 | int err; |
| 2647 | |
| 2648 | switch (advice) { |
| 2649 | case MADV_MERGEABLE: |
| 2650 | if (vma->vm_flags & VM_MERGEABLE) |
| 2651 | return 0; |
| 2652 | if (!vma_ksm_compatible(vma)) |
| 2653 | return 0; |
| 2654 | |
| 2655 | if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) { |
| 2656 | err = __ksm_enter(mm); |
| 2657 | if (err) |
| 2658 | return err; |
| 2659 | } |
| 2660 | |
| 2661 | *vm_flags |= VM_MERGEABLE; |
| 2662 | break; |
| 2663 | |
| 2664 | case MADV_UNMERGEABLE: |
| 2665 | if (!(*vm_flags & VM_MERGEABLE)) |
| 2666 | return 0; /* just ignore the advice */ |
| 2667 | |
| 2668 | if (vma->anon_vma) { |
| 2669 | err = unmerge_ksm_pages(vma, start, end); |
| 2670 | if (err) |
| 2671 | return err; |
| 2672 | } |
| 2673 | |
| 2674 | *vm_flags &= ~VM_MERGEABLE; |
| 2675 | break; |
| 2676 | } |
| 2677 | |
| 2678 | return 0; |
| 2679 | } |
| 2680 | EXPORT_SYMBOL_GPL(ksm_madvise); |
| 2681 | |
| 2682 | int __ksm_enter(struct mm_struct *mm) |
| 2683 | { |
| 2684 | struct ksm_mm_slot *mm_slot; |
| 2685 | struct mm_slot *slot; |
| 2686 | int needs_wakeup; |
| 2687 | |
| 2688 | mm_slot = mm_slot_alloc(mm_slot_cache); |
| 2689 | if (!mm_slot) |
| 2690 | return -ENOMEM; |
| 2691 | |
| 2692 | slot = &mm_slot->slot; |
| 2693 | |
| 2694 | /* Check ksm_run too? Would need tighter locking */ |
| 2695 | needs_wakeup = list_empty(&ksm_mm_head.slot.mm_node); |
| 2696 | |
| 2697 | spin_lock(&ksm_mmlist_lock); |
| 2698 | mm_slot_insert(mm_slots_hash, mm, slot); |
| 2699 | /* |
| 2700 | * When KSM_RUN_MERGE (or KSM_RUN_STOP), |
| 2701 | * insert just behind the scanning cursor, to let the area settle |
| 2702 | * down a little; when fork is followed by immediate exec, we don't |
| 2703 | * want ksmd to waste time setting up and tearing down an rmap_list. |
| 2704 | * |
| 2705 | * But when KSM_RUN_UNMERGE, it's important to insert ahead of its |
| 2706 | * scanning cursor, otherwise KSM pages in newly forked mms will be |
| 2707 | * missed: then we might as well insert at the end of the list. |
| 2708 | */ |
| 2709 | if (ksm_run & KSM_RUN_UNMERGE) |
| 2710 | list_add_tail(&slot->mm_node, &ksm_mm_head.slot.mm_node); |
| 2711 | else |
| 2712 | list_add_tail(&slot->mm_node, &ksm_scan.mm_slot->slot.mm_node); |
| 2713 | spin_unlock(&ksm_mmlist_lock); |
| 2714 | |
| 2715 | set_bit(MMF_VM_MERGEABLE, &mm->flags); |
| 2716 | mmgrab(mm); |
| 2717 | |
| 2718 | if (needs_wakeup) |
| 2719 | wake_up_interruptible(&ksm_thread_wait); |
| 2720 | |
| 2721 | trace_ksm_enter(mm); |
| 2722 | return 0; |
| 2723 | } |
| 2724 | |
| 2725 | void __ksm_exit(struct mm_struct *mm) |
| 2726 | { |
| 2727 | struct ksm_mm_slot *mm_slot; |
| 2728 | struct mm_slot *slot; |
| 2729 | int easy_to_free = 0; |
| 2730 | |
| 2731 | /* |
| 2732 | * This process is exiting: if it's straightforward (as is the |
| 2733 | * case when ksmd was never running), free mm_slot immediately. |
| 2734 | * But if it's at the cursor or has rmap_items linked to it, use |
| 2735 | * mmap_lock to synchronize with any break_cows before pagetables |
| 2736 | * are freed, and leave the mm_slot on the list for ksmd to free. |
| 2737 | * Beware: ksm may already have noticed it exiting and freed the slot. |
| 2738 | */ |
| 2739 | |
| 2740 | spin_lock(&ksm_mmlist_lock); |
| 2741 | slot = mm_slot_lookup(mm_slots_hash, mm); |
| 2742 | mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); |
| 2743 | if (mm_slot && ksm_scan.mm_slot != mm_slot) { |
| 2744 | if (!mm_slot->rmap_list) { |
| 2745 | hash_del(&slot->hash); |
| 2746 | list_del(&slot->mm_node); |
| 2747 | easy_to_free = 1; |
| 2748 | } else { |
| 2749 | list_move(&slot->mm_node, |
| 2750 | &ksm_scan.mm_slot->slot.mm_node); |
| 2751 | } |
| 2752 | } |
| 2753 | spin_unlock(&ksm_mmlist_lock); |
| 2754 | |
| 2755 | if (easy_to_free) { |
| 2756 | mm_slot_free(mm_slot_cache, mm_slot); |
| 2757 | clear_bit(MMF_VM_MERGE_ANY, &mm->flags); |
| 2758 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); |
| 2759 | mmdrop(mm); |
| 2760 | } else if (mm_slot) { |
| 2761 | mmap_write_lock(mm); |
| 2762 | mmap_write_unlock(mm); |
| 2763 | } |
| 2764 | |
| 2765 | trace_ksm_exit(mm); |
| 2766 | } |
| 2767 | |
| 2768 | struct page *ksm_might_need_to_copy(struct page *page, |
| 2769 | struct vm_area_struct *vma, unsigned long address) |
| 2770 | { |
| 2771 | struct folio *folio = page_folio(page); |
| 2772 | struct anon_vma *anon_vma = folio_anon_vma(folio); |
| 2773 | struct page *new_page; |
| 2774 | |
| 2775 | if (PageKsm(page)) { |
| 2776 | if (page_stable_node(page) && |
| 2777 | !(ksm_run & KSM_RUN_UNMERGE)) |
| 2778 | return page; /* no need to copy it */ |
| 2779 | } else if (!anon_vma) { |
| 2780 | return page; /* no need to copy it */ |
| 2781 | } else if (page->index == linear_page_index(vma, address) && |
| 2782 | anon_vma->root == vma->anon_vma->root) { |
| 2783 | return page; /* still no need to copy it */ |
| 2784 | } |
| 2785 | if (!PageUptodate(page)) |
| 2786 | return page; /* let do_swap_page report the error */ |
| 2787 | |
| 2788 | new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); |
| 2789 | if (new_page && |
| 2790 | mem_cgroup_charge(page_folio(new_page), vma->vm_mm, GFP_KERNEL)) { |
| 2791 | put_page(new_page); |
| 2792 | new_page = NULL; |
| 2793 | } |
| 2794 | if (new_page) { |
| 2795 | if (copy_mc_user_highpage(new_page, page, address, vma)) { |
| 2796 | put_page(new_page); |
| 2797 | memory_failure_queue(page_to_pfn(page), 0); |
| 2798 | return ERR_PTR(-EHWPOISON); |
| 2799 | } |
| 2800 | SetPageDirty(new_page); |
| 2801 | __SetPageUptodate(new_page); |
| 2802 | __SetPageLocked(new_page); |
| 2803 | #ifdef CONFIG_SWAP |
| 2804 | count_vm_event(KSM_SWPIN_COPY); |
| 2805 | #endif |
| 2806 | } |
| 2807 | |
| 2808 | return new_page; |
| 2809 | } |
| 2810 | |
| 2811 | void rmap_walk_ksm(struct folio *folio, struct rmap_walk_control *rwc) |
| 2812 | { |
| 2813 | struct ksm_stable_node *stable_node; |
| 2814 | struct ksm_rmap_item *rmap_item; |
| 2815 | int search_new_forks = 0; |
| 2816 | |
| 2817 | VM_BUG_ON_FOLIO(!folio_test_ksm(folio), folio); |
| 2818 | |
| 2819 | /* |
| 2820 | * Rely on the page lock to protect against concurrent modifications |
| 2821 | * to that page's node of the stable tree. |
| 2822 | */ |
| 2823 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); |
| 2824 | |
| 2825 | stable_node = folio_stable_node(folio); |
| 2826 | if (!stable_node) |
| 2827 | return; |
| 2828 | again: |
| 2829 | hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) { |
| 2830 | struct anon_vma *anon_vma = rmap_item->anon_vma; |
| 2831 | struct anon_vma_chain *vmac; |
| 2832 | struct vm_area_struct *vma; |
| 2833 | |
| 2834 | cond_resched(); |
| 2835 | if (!anon_vma_trylock_read(anon_vma)) { |
| 2836 | if (rwc->try_lock) { |
| 2837 | rwc->contended = true; |
| 2838 | return; |
| 2839 | } |
| 2840 | anon_vma_lock_read(anon_vma); |
| 2841 | } |
| 2842 | anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root, |
| 2843 | 0, ULONG_MAX) { |
| 2844 | unsigned long addr; |
| 2845 | |
| 2846 | cond_resched(); |
| 2847 | vma = vmac->vma; |
| 2848 | |
| 2849 | /* Ignore the stable/unstable/sqnr flags */ |
| 2850 | addr = rmap_item->address & PAGE_MASK; |
| 2851 | |
| 2852 | if (addr < vma->vm_start || addr >= vma->vm_end) |
| 2853 | continue; |
| 2854 | /* |
| 2855 | * Initially we examine only the vma which covers this |
| 2856 | * rmap_item; but later, if there is still work to do, |
| 2857 | * we examine covering vmas in other mms: in case they |
| 2858 | * were forked from the original since ksmd passed. |
| 2859 | */ |
| 2860 | if ((rmap_item->mm == vma->vm_mm) == search_new_forks) |
| 2861 | continue; |
| 2862 | |
| 2863 | if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) |
| 2864 | continue; |
| 2865 | |
| 2866 | if (!rwc->rmap_one(folio, vma, addr, rwc->arg)) { |
| 2867 | anon_vma_unlock_read(anon_vma); |
| 2868 | return; |
| 2869 | } |
| 2870 | if (rwc->done && rwc->done(folio)) { |
| 2871 | anon_vma_unlock_read(anon_vma); |
| 2872 | return; |
| 2873 | } |
| 2874 | } |
| 2875 | anon_vma_unlock_read(anon_vma); |
| 2876 | } |
| 2877 | if (!search_new_forks++) |
| 2878 | goto again; |
| 2879 | } |
| 2880 | |
| 2881 | #ifdef CONFIG_MEMORY_FAILURE |
| 2882 | /* |
| 2883 | * Collect processes when the error hit an ksm page. |
| 2884 | */ |
| 2885 | void collect_procs_ksm(struct page *page, struct list_head *to_kill, |
| 2886 | int force_early) |
| 2887 | { |
| 2888 | struct ksm_stable_node *stable_node; |
| 2889 | struct ksm_rmap_item *rmap_item; |
| 2890 | struct folio *folio = page_folio(page); |
| 2891 | struct vm_area_struct *vma; |
| 2892 | struct task_struct *tsk; |
| 2893 | |
| 2894 | stable_node = folio_stable_node(folio); |
| 2895 | if (!stable_node) |
| 2896 | return; |
| 2897 | hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) { |
| 2898 | struct anon_vma *av = rmap_item->anon_vma; |
| 2899 | |
| 2900 | anon_vma_lock_read(av); |
| 2901 | read_lock(&tasklist_lock); |
| 2902 | for_each_process(tsk) { |
| 2903 | struct anon_vma_chain *vmac; |
| 2904 | unsigned long addr; |
| 2905 | struct task_struct *t = |
| 2906 | task_early_kill(tsk, force_early); |
| 2907 | if (!t) |
| 2908 | continue; |
| 2909 | anon_vma_interval_tree_foreach(vmac, &av->rb_root, 0, |
| 2910 | ULONG_MAX) |
| 2911 | { |
| 2912 | vma = vmac->vma; |
| 2913 | if (vma->vm_mm == t->mm) { |
| 2914 | addr = rmap_item->address & PAGE_MASK; |
| 2915 | add_to_kill_ksm(t, page, vma, to_kill, |
| 2916 | addr); |
| 2917 | } |
| 2918 | } |
| 2919 | } |
| 2920 | read_unlock(&tasklist_lock); |
| 2921 | anon_vma_unlock_read(av); |
| 2922 | } |
| 2923 | } |
| 2924 | #endif |
| 2925 | |
| 2926 | #ifdef CONFIG_MIGRATION |
| 2927 | void folio_migrate_ksm(struct folio *newfolio, struct folio *folio) |
| 2928 | { |
| 2929 | struct ksm_stable_node *stable_node; |
| 2930 | |
| 2931 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); |
| 2932 | VM_BUG_ON_FOLIO(!folio_test_locked(newfolio), newfolio); |
| 2933 | VM_BUG_ON_FOLIO(newfolio->mapping != folio->mapping, newfolio); |
| 2934 | |
| 2935 | stable_node = folio_stable_node(folio); |
| 2936 | if (stable_node) { |
| 2937 | VM_BUG_ON_FOLIO(stable_node->kpfn != folio_pfn(folio), folio); |
| 2938 | stable_node->kpfn = folio_pfn(newfolio); |
| 2939 | /* |
| 2940 | * newfolio->mapping was set in advance; now we need smp_wmb() |
| 2941 | * to make sure that the new stable_node->kpfn is visible |
| 2942 | * to get_ksm_page() before it can see that folio->mapping |
| 2943 | * has gone stale (or that folio_test_swapcache has been cleared). |
| 2944 | */ |
| 2945 | smp_wmb(); |
| 2946 | set_page_stable_node(&folio->page, NULL); |
| 2947 | } |
| 2948 | } |
| 2949 | #endif /* CONFIG_MIGRATION */ |
| 2950 | |
| 2951 | #ifdef CONFIG_MEMORY_HOTREMOVE |
| 2952 | static void wait_while_offlining(void) |
| 2953 | { |
| 2954 | while (ksm_run & KSM_RUN_OFFLINE) { |
| 2955 | mutex_unlock(&ksm_thread_mutex); |
| 2956 | wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE), |
| 2957 | TASK_UNINTERRUPTIBLE); |
| 2958 | mutex_lock(&ksm_thread_mutex); |
| 2959 | } |
| 2960 | } |
| 2961 | |
| 2962 | static bool stable_node_dup_remove_range(struct ksm_stable_node *stable_node, |
| 2963 | unsigned long start_pfn, |
| 2964 | unsigned long end_pfn) |
| 2965 | { |
| 2966 | if (stable_node->kpfn >= start_pfn && |
| 2967 | stable_node->kpfn < end_pfn) { |
| 2968 | /* |
| 2969 | * Don't get_ksm_page, page has already gone: |
| 2970 | * which is why we keep kpfn instead of page* |
| 2971 | */ |
| 2972 | remove_node_from_stable_tree(stable_node); |
| 2973 | return true; |
| 2974 | } |
| 2975 | return false; |
| 2976 | } |
| 2977 | |
| 2978 | static bool stable_node_chain_remove_range(struct ksm_stable_node *stable_node, |
| 2979 | unsigned long start_pfn, |
| 2980 | unsigned long end_pfn, |
| 2981 | struct rb_root *root) |
| 2982 | { |
| 2983 | struct ksm_stable_node *dup; |
| 2984 | struct hlist_node *hlist_safe; |
| 2985 | |
| 2986 | if (!is_stable_node_chain(stable_node)) { |
| 2987 | VM_BUG_ON(is_stable_node_dup(stable_node)); |
| 2988 | return stable_node_dup_remove_range(stable_node, start_pfn, |
| 2989 | end_pfn); |
| 2990 | } |
| 2991 | |
| 2992 | hlist_for_each_entry_safe(dup, hlist_safe, |
| 2993 | &stable_node->hlist, hlist_dup) { |
| 2994 | VM_BUG_ON(!is_stable_node_dup(dup)); |
| 2995 | stable_node_dup_remove_range(dup, start_pfn, end_pfn); |
| 2996 | } |
| 2997 | if (hlist_empty(&stable_node->hlist)) { |
| 2998 | free_stable_node_chain(stable_node, root); |
| 2999 | return true; /* notify caller that tree was rebalanced */ |
| 3000 | } else |
| 3001 | return false; |
| 3002 | } |
| 3003 | |
| 3004 | static void ksm_check_stable_tree(unsigned long start_pfn, |
| 3005 | unsigned long end_pfn) |
| 3006 | { |
| 3007 | struct ksm_stable_node *stable_node, *next; |
| 3008 | struct rb_node *node; |
| 3009 | int nid; |
| 3010 | |
| 3011 | for (nid = 0; nid < ksm_nr_node_ids; nid++) { |
| 3012 | node = rb_first(root_stable_tree + nid); |
| 3013 | while (node) { |
| 3014 | stable_node = rb_entry(node, struct ksm_stable_node, node); |
| 3015 | if (stable_node_chain_remove_range(stable_node, |
| 3016 | start_pfn, end_pfn, |
| 3017 | root_stable_tree + |
| 3018 | nid)) |
| 3019 | node = rb_first(root_stable_tree + nid); |
| 3020 | else |
| 3021 | node = rb_next(node); |
| 3022 | cond_resched(); |
| 3023 | } |
| 3024 | } |
| 3025 | list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) { |
| 3026 | if (stable_node->kpfn >= start_pfn && |
| 3027 | stable_node->kpfn < end_pfn) |
| 3028 | remove_node_from_stable_tree(stable_node); |
| 3029 | cond_resched(); |
| 3030 | } |
| 3031 | } |
| 3032 | |
| 3033 | static int ksm_memory_callback(struct notifier_block *self, |
| 3034 | unsigned long action, void *arg) |
| 3035 | { |
| 3036 | struct memory_notify *mn = arg; |
| 3037 | |
| 3038 | switch (action) { |
| 3039 | case MEM_GOING_OFFLINE: |
| 3040 | /* |
| 3041 | * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items() |
| 3042 | * and remove_all_stable_nodes() while memory is going offline: |
| 3043 | * it is unsafe for them to touch the stable tree at this time. |
| 3044 | * But unmerge_ksm_pages(), rmap lookups and other entry points |
| 3045 | * which do not need the ksm_thread_mutex are all safe. |
| 3046 | */ |
| 3047 | mutex_lock(&ksm_thread_mutex); |
| 3048 | ksm_run |= KSM_RUN_OFFLINE; |
| 3049 | mutex_unlock(&ksm_thread_mutex); |
| 3050 | break; |
| 3051 | |
| 3052 | case MEM_OFFLINE: |
| 3053 | /* |
| 3054 | * Most of the work is done by page migration; but there might |
| 3055 | * be a few stable_nodes left over, still pointing to struct |
| 3056 | * pages which have been offlined: prune those from the tree, |
| 3057 | * otherwise get_ksm_page() might later try to access a |
| 3058 | * non-existent struct page. |
| 3059 | */ |
| 3060 | ksm_check_stable_tree(mn->start_pfn, |
| 3061 | mn->start_pfn + mn->nr_pages); |
| 3062 | fallthrough; |
| 3063 | case MEM_CANCEL_OFFLINE: |
| 3064 | mutex_lock(&ksm_thread_mutex); |
| 3065 | ksm_run &= ~KSM_RUN_OFFLINE; |
| 3066 | mutex_unlock(&ksm_thread_mutex); |
| 3067 | |
| 3068 | smp_mb(); /* wake_up_bit advises this */ |
| 3069 | wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE)); |
| 3070 | break; |
| 3071 | } |
| 3072 | return NOTIFY_OK; |
| 3073 | } |
| 3074 | #else |
| 3075 | static void wait_while_offlining(void) |
| 3076 | { |
| 3077 | } |
| 3078 | #endif /* CONFIG_MEMORY_HOTREMOVE */ |
| 3079 | |
| 3080 | #ifdef CONFIG_PROC_FS |
| 3081 | long ksm_process_profit(struct mm_struct *mm) |
| 3082 | { |
| 3083 | return mm->ksm_merging_pages * PAGE_SIZE - |
| 3084 | mm->ksm_rmap_items * sizeof(struct ksm_rmap_item); |
| 3085 | } |
| 3086 | #endif /* CONFIG_PROC_FS */ |
| 3087 | |
| 3088 | #ifdef CONFIG_SYSFS |
| 3089 | /* |
| 3090 | * This all compiles without CONFIG_SYSFS, but is a waste of space. |
| 3091 | */ |
| 3092 | |
| 3093 | #define KSM_ATTR_RO(_name) \ |
| 3094 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) |
| 3095 | #define KSM_ATTR(_name) \ |
| 3096 | static struct kobj_attribute _name##_attr = __ATTR_RW(_name) |
| 3097 | |
| 3098 | static ssize_t sleep_millisecs_show(struct kobject *kobj, |
| 3099 | struct kobj_attribute *attr, char *buf) |
| 3100 | { |
| 3101 | return sysfs_emit(buf, "%u\n", ksm_thread_sleep_millisecs); |
| 3102 | } |
| 3103 | |
| 3104 | static ssize_t sleep_millisecs_store(struct kobject *kobj, |
| 3105 | struct kobj_attribute *attr, |
| 3106 | const char *buf, size_t count) |
| 3107 | { |
| 3108 | unsigned int msecs; |
| 3109 | int err; |
| 3110 | |
| 3111 | err = kstrtouint(buf, 10, &msecs); |
| 3112 | if (err) |
| 3113 | return -EINVAL; |
| 3114 | |
| 3115 | ksm_thread_sleep_millisecs = msecs; |
| 3116 | wake_up_interruptible(&ksm_iter_wait); |
| 3117 | |
| 3118 | return count; |
| 3119 | } |
| 3120 | KSM_ATTR(sleep_millisecs); |
| 3121 | |
| 3122 | static ssize_t pages_to_scan_show(struct kobject *kobj, |
| 3123 | struct kobj_attribute *attr, char *buf) |
| 3124 | { |
| 3125 | return sysfs_emit(buf, "%u\n", ksm_thread_pages_to_scan); |
| 3126 | } |
| 3127 | |
| 3128 | static ssize_t pages_to_scan_store(struct kobject *kobj, |
| 3129 | struct kobj_attribute *attr, |
| 3130 | const char *buf, size_t count) |
| 3131 | { |
| 3132 | unsigned int nr_pages; |
| 3133 | int err; |
| 3134 | |
| 3135 | err = kstrtouint(buf, 10, &nr_pages); |
| 3136 | if (err) |
| 3137 | return -EINVAL; |
| 3138 | |
| 3139 | ksm_thread_pages_to_scan = nr_pages; |
| 3140 | |
| 3141 | return count; |
| 3142 | } |
| 3143 | KSM_ATTR(pages_to_scan); |
| 3144 | |
| 3145 | static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr, |
| 3146 | char *buf) |
| 3147 | { |
| 3148 | return sysfs_emit(buf, "%lu\n", ksm_run); |
| 3149 | } |
| 3150 | |
| 3151 | static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr, |
| 3152 | const char *buf, size_t count) |
| 3153 | { |
| 3154 | unsigned int flags; |
| 3155 | int err; |
| 3156 | |
| 3157 | err = kstrtouint(buf, 10, &flags); |
| 3158 | if (err) |
| 3159 | return -EINVAL; |
| 3160 | if (flags > KSM_RUN_UNMERGE) |
| 3161 | return -EINVAL; |
| 3162 | |
| 3163 | /* |
| 3164 | * KSM_RUN_MERGE sets ksmd running, and 0 stops it running. |
| 3165 | * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items, |
| 3166 | * breaking COW to free the pages_shared (but leaves mm_slots |
| 3167 | * on the list for when ksmd may be set running again). |
| 3168 | */ |
| 3169 | |
| 3170 | mutex_lock(&ksm_thread_mutex); |
| 3171 | wait_while_offlining(); |
| 3172 | if (ksm_run != flags) { |
| 3173 | ksm_run = flags; |
| 3174 | if (flags & KSM_RUN_UNMERGE) { |
| 3175 | set_current_oom_origin(); |
| 3176 | err = unmerge_and_remove_all_rmap_items(); |
| 3177 | clear_current_oom_origin(); |
| 3178 | if (err) { |
| 3179 | ksm_run = KSM_RUN_STOP; |
| 3180 | count = err; |
| 3181 | } |
| 3182 | } |
| 3183 | } |
| 3184 | mutex_unlock(&ksm_thread_mutex); |
| 3185 | |
| 3186 | if (flags & KSM_RUN_MERGE) |
| 3187 | wake_up_interruptible(&ksm_thread_wait); |
| 3188 | |
| 3189 | return count; |
| 3190 | } |
| 3191 | KSM_ATTR(run); |
| 3192 | |
| 3193 | #ifdef CONFIG_NUMA |
| 3194 | static ssize_t merge_across_nodes_show(struct kobject *kobj, |
| 3195 | struct kobj_attribute *attr, char *buf) |
| 3196 | { |
| 3197 | return sysfs_emit(buf, "%u\n", ksm_merge_across_nodes); |
| 3198 | } |
| 3199 | |
| 3200 | static ssize_t merge_across_nodes_store(struct kobject *kobj, |
| 3201 | struct kobj_attribute *attr, |
| 3202 | const char *buf, size_t count) |
| 3203 | { |
| 3204 | int err; |
| 3205 | unsigned long knob; |
| 3206 | |
| 3207 | err = kstrtoul(buf, 10, &knob); |
| 3208 | if (err) |
| 3209 | return err; |
| 3210 | if (knob > 1) |
| 3211 | return -EINVAL; |
| 3212 | |
| 3213 | mutex_lock(&ksm_thread_mutex); |
| 3214 | wait_while_offlining(); |
| 3215 | if (ksm_merge_across_nodes != knob) { |
| 3216 | if (ksm_pages_shared || remove_all_stable_nodes()) |
| 3217 | err = -EBUSY; |
| 3218 | else if (root_stable_tree == one_stable_tree) { |
| 3219 | struct rb_root *buf; |
| 3220 | /* |
| 3221 | * This is the first time that we switch away from the |
| 3222 | * default of merging across nodes: must now allocate |
| 3223 | * a buffer to hold as many roots as may be needed. |
| 3224 | * Allocate stable and unstable together: |
| 3225 | * MAXSMP NODES_SHIFT 10 will use 16kB. |
| 3226 | */ |
| 3227 | buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf), |
| 3228 | GFP_KERNEL); |
| 3229 | /* Let us assume that RB_ROOT is NULL is zero */ |
| 3230 | if (!buf) |
| 3231 | err = -ENOMEM; |
| 3232 | else { |
| 3233 | root_stable_tree = buf; |
| 3234 | root_unstable_tree = buf + nr_node_ids; |
| 3235 | /* Stable tree is empty but not the unstable */ |
| 3236 | root_unstable_tree[0] = one_unstable_tree[0]; |
| 3237 | } |
| 3238 | } |
| 3239 | if (!err) { |
| 3240 | ksm_merge_across_nodes = knob; |
| 3241 | ksm_nr_node_ids = knob ? 1 : nr_node_ids; |
| 3242 | } |
| 3243 | } |
| 3244 | mutex_unlock(&ksm_thread_mutex); |
| 3245 | |
| 3246 | return err ? err : count; |
| 3247 | } |
| 3248 | KSM_ATTR(merge_across_nodes); |
| 3249 | #endif |
| 3250 | |
| 3251 | static ssize_t use_zero_pages_show(struct kobject *kobj, |
| 3252 | struct kobj_attribute *attr, char *buf) |
| 3253 | { |
| 3254 | return sysfs_emit(buf, "%u\n", ksm_use_zero_pages); |
| 3255 | } |
| 3256 | static ssize_t use_zero_pages_store(struct kobject *kobj, |
| 3257 | struct kobj_attribute *attr, |
| 3258 | const char *buf, size_t count) |
| 3259 | { |
| 3260 | int err; |
| 3261 | bool value; |
| 3262 | |
| 3263 | err = kstrtobool(buf, &value); |
| 3264 | if (err) |
| 3265 | return -EINVAL; |
| 3266 | |
| 3267 | ksm_use_zero_pages = value; |
| 3268 | |
| 3269 | return count; |
| 3270 | } |
| 3271 | KSM_ATTR(use_zero_pages); |
| 3272 | |
| 3273 | static ssize_t max_page_sharing_show(struct kobject *kobj, |
| 3274 | struct kobj_attribute *attr, char *buf) |
| 3275 | { |
| 3276 | return sysfs_emit(buf, "%u\n", ksm_max_page_sharing); |
| 3277 | } |
| 3278 | |
| 3279 | static ssize_t max_page_sharing_store(struct kobject *kobj, |
| 3280 | struct kobj_attribute *attr, |
| 3281 | const char *buf, size_t count) |
| 3282 | { |
| 3283 | int err; |
| 3284 | int knob; |
| 3285 | |
| 3286 | err = kstrtoint(buf, 10, &knob); |
| 3287 | if (err) |
| 3288 | return err; |
| 3289 | /* |
| 3290 | * When a KSM page is created it is shared by 2 mappings. This |
| 3291 | * being a signed comparison, it implicitly verifies it's not |
| 3292 | * negative. |
| 3293 | */ |
| 3294 | if (knob < 2) |
| 3295 | return -EINVAL; |
| 3296 | |
| 3297 | if (READ_ONCE(ksm_max_page_sharing) == knob) |
| 3298 | return count; |
| 3299 | |
| 3300 | mutex_lock(&ksm_thread_mutex); |
| 3301 | wait_while_offlining(); |
| 3302 | if (ksm_max_page_sharing != knob) { |
| 3303 | if (ksm_pages_shared || remove_all_stable_nodes()) |
| 3304 | err = -EBUSY; |
| 3305 | else |
| 3306 | ksm_max_page_sharing = knob; |
| 3307 | } |
| 3308 | mutex_unlock(&ksm_thread_mutex); |
| 3309 | |
| 3310 | return err ? err : count; |
| 3311 | } |
| 3312 | KSM_ATTR(max_page_sharing); |
| 3313 | |
| 3314 | static ssize_t pages_shared_show(struct kobject *kobj, |
| 3315 | struct kobj_attribute *attr, char *buf) |
| 3316 | { |
| 3317 | return sysfs_emit(buf, "%lu\n", ksm_pages_shared); |
| 3318 | } |
| 3319 | KSM_ATTR_RO(pages_shared); |
| 3320 | |
| 3321 | static ssize_t pages_sharing_show(struct kobject *kobj, |
| 3322 | struct kobj_attribute *attr, char *buf) |
| 3323 | { |
| 3324 | return sysfs_emit(buf, "%lu\n", ksm_pages_sharing); |
| 3325 | } |
| 3326 | KSM_ATTR_RO(pages_sharing); |
| 3327 | |
| 3328 | static ssize_t pages_unshared_show(struct kobject *kobj, |
| 3329 | struct kobj_attribute *attr, char *buf) |
| 3330 | { |
| 3331 | return sysfs_emit(buf, "%lu\n", ksm_pages_unshared); |
| 3332 | } |
| 3333 | KSM_ATTR_RO(pages_unshared); |
| 3334 | |
| 3335 | static ssize_t pages_volatile_show(struct kobject *kobj, |
| 3336 | struct kobj_attribute *attr, char *buf) |
| 3337 | { |
| 3338 | long ksm_pages_volatile; |
| 3339 | |
| 3340 | ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared |
| 3341 | - ksm_pages_sharing - ksm_pages_unshared; |
| 3342 | /* |
| 3343 | * It was not worth any locking to calculate that statistic, |
| 3344 | * but it might therefore sometimes be negative: conceal that. |
| 3345 | */ |
| 3346 | if (ksm_pages_volatile < 0) |
| 3347 | ksm_pages_volatile = 0; |
| 3348 | return sysfs_emit(buf, "%ld\n", ksm_pages_volatile); |
| 3349 | } |
| 3350 | KSM_ATTR_RO(pages_volatile); |
| 3351 | |
| 3352 | static ssize_t general_profit_show(struct kobject *kobj, |
| 3353 | struct kobj_attribute *attr, char *buf) |
| 3354 | { |
| 3355 | long general_profit; |
| 3356 | |
| 3357 | general_profit = ksm_pages_sharing * PAGE_SIZE - |
| 3358 | ksm_rmap_items * sizeof(struct ksm_rmap_item); |
| 3359 | |
| 3360 | return sysfs_emit(buf, "%ld\n", general_profit); |
| 3361 | } |
| 3362 | KSM_ATTR_RO(general_profit); |
| 3363 | |
| 3364 | static ssize_t stable_node_dups_show(struct kobject *kobj, |
| 3365 | struct kobj_attribute *attr, char *buf) |
| 3366 | { |
| 3367 | return sysfs_emit(buf, "%lu\n", ksm_stable_node_dups); |
| 3368 | } |
| 3369 | KSM_ATTR_RO(stable_node_dups); |
| 3370 | |
| 3371 | static ssize_t stable_node_chains_show(struct kobject *kobj, |
| 3372 | struct kobj_attribute *attr, char *buf) |
| 3373 | { |
| 3374 | return sysfs_emit(buf, "%lu\n", ksm_stable_node_chains); |
| 3375 | } |
| 3376 | KSM_ATTR_RO(stable_node_chains); |
| 3377 | |
| 3378 | static ssize_t |
| 3379 | stable_node_chains_prune_millisecs_show(struct kobject *kobj, |
| 3380 | struct kobj_attribute *attr, |
| 3381 | char *buf) |
| 3382 | { |
| 3383 | return sysfs_emit(buf, "%u\n", ksm_stable_node_chains_prune_millisecs); |
| 3384 | } |
| 3385 | |
| 3386 | static ssize_t |
| 3387 | stable_node_chains_prune_millisecs_store(struct kobject *kobj, |
| 3388 | struct kobj_attribute *attr, |
| 3389 | const char *buf, size_t count) |
| 3390 | { |
| 3391 | unsigned int msecs; |
| 3392 | int err; |
| 3393 | |
| 3394 | err = kstrtouint(buf, 10, &msecs); |
| 3395 | if (err) |
| 3396 | return -EINVAL; |
| 3397 | |
| 3398 | ksm_stable_node_chains_prune_millisecs = msecs; |
| 3399 | |
| 3400 | return count; |
| 3401 | } |
| 3402 | KSM_ATTR(stable_node_chains_prune_millisecs); |
| 3403 | |
| 3404 | static ssize_t full_scans_show(struct kobject *kobj, |
| 3405 | struct kobj_attribute *attr, char *buf) |
| 3406 | { |
| 3407 | return sysfs_emit(buf, "%lu\n", ksm_scan.seqnr); |
| 3408 | } |
| 3409 | KSM_ATTR_RO(full_scans); |
| 3410 | |
| 3411 | static struct attribute *ksm_attrs[] = { |
| 3412 | &sleep_millisecs_attr.attr, |
| 3413 | &pages_to_scan_attr.attr, |
| 3414 | &run_attr.attr, |
| 3415 | &pages_shared_attr.attr, |
| 3416 | &pages_sharing_attr.attr, |
| 3417 | &pages_unshared_attr.attr, |
| 3418 | &pages_volatile_attr.attr, |
| 3419 | &full_scans_attr.attr, |
| 3420 | #ifdef CONFIG_NUMA |
| 3421 | &merge_across_nodes_attr.attr, |
| 3422 | #endif |
| 3423 | &max_page_sharing_attr.attr, |
| 3424 | &stable_node_chains_attr.attr, |
| 3425 | &stable_node_dups_attr.attr, |
| 3426 | &stable_node_chains_prune_millisecs_attr.attr, |
| 3427 | &use_zero_pages_attr.attr, |
| 3428 | &general_profit_attr.attr, |
| 3429 | NULL, |
| 3430 | }; |
| 3431 | |
| 3432 | static const struct attribute_group ksm_attr_group = { |
| 3433 | .attrs = ksm_attrs, |
| 3434 | .name = "ksm", |
| 3435 | }; |
| 3436 | #endif /* CONFIG_SYSFS */ |
| 3437 | |
| 3438 | static int __init ksm_init(void) |
| 3439 | { |
| 3440 | struct task_struct *ksm_thread; |
| 3441 | int err; |
| 3442 | |
| 3443 | /* The correct value depends on page size and endianness */ |
| 3444 | zero_checksum = calc_checksum(ZERO_PAGE(0)); |
| 3445 | /* Default to false for backwards compatibility */ |
| 3446 | ksm_use_zero_pages = false; |
| 3447 | |
| 3448 | err = ksm_slab_init(); |
| 3449 | if (err) |
| 3450 | goto out; |
| 3451 | |
| 3452 | ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd"); |
| 3453 | if (IS_ERR(ksm_thread)) { |
| 3454 | pr_err("ksm: creating kthread failed\n"); |
| 3455 | err = PTR_ERR(ksm_thread); |
| 3456 | goto out_free; |
| 3457 | } |
| 3458 | |
| 3459 | #ifdef CONFIG_SYSFS |
| 3460 | err = sysfs_create_group(mm_kobj, &ksm_attr_group); |
| 3461 | if (err) { |
| 3462 | pr_err("ksm: register sysfs failed\n"); |
| 3463 | kthread_stop(ksm_thread); |
| 3464 | goto out_free; |
| 3465 | } |
| 3466 | #else |
| 3467 | ksm_run = KSM_RUN_MERGE; /* no way for user to start it */ |
| 3468 | |
| 3469 | #endif /* CONFIG_SYSFS */ |
| 3470 | |
| 3471 | #ifdef CONFIG_MEMORY_HOTREMOVE |
| 3472 | /* There is no significance to this priority 100 */ |
| 3473 | hotplug_memory_notifier(ksm_memory_callback, KSM_CALLBACK_PRI); |
| 3474 | #endif |
| 3475 | return 0; |
| 3476 | |
| 3477 | out_free: |
| 3478 | ksm_slab_free(); |
| 3479 | out: |
| 3480 | return err; |
| 3481 | } |
| 3482 | subsys_initcall(ksm_init); |