1.. _list_rcu_doc: 2 3Using RCU to Protect Read-Mostly Linked Lists 4============================================= 5 6One of the most common uses of RCU is protecting read-mostly linked lists 7(``struct list_head`` in list.h). One big advantage of this approach is 8that all of the required memory ordering is provided by the list macros. 9This document describes several list-based RCU use cases. 10 11When iterating a list while holding the rcu_read_lock(), writers may 12modify the list. The reader is guaranteed to see all of the elements 13which were added to the list before they acquired the rcu_read_lock() 14and are still on the list when they drop the rcu_read_unlock(). 15Elements which are added to, or removed from the list may or may not 16be seen. If the writer calls list_replace_rcu(), the reader may see 17either the old element or the new element; they will not see both, 18nor will they see neither. 19 20 21Example 1: Read-mostly list: Deferred Destruction 22------------------------------------------------- 23 24A widely used usecase for RCU lists in the kernel is lockless iteration over 25all processes in the system. ``task_struct::tasks`` represents the list node that 26links all the processes. The list can be traversed in parallel to any list 27additions or removals. 28 29The traversal of the list is done using ``for_each_process()`` which is defined 30by the 2 macros:: 31 32 #define next_task(p) \ 33 list_entry_rcu((p)->tasks.next, struct task_struct, tasks) 34 35 #define for_each_process(p) \ 36 for (p = &init_task ; (p = next_task(p)) != &init_task ; ) 37 38The code traversing the list of all processes typically looks like:: 39 40 rcu_read_lock(); 41 for_each_process(p) { 42 /* Do something with p */ 43 } 44 rcu_read_unlock(); 45 46The simplified and heavily inlined code for removing a process from a 47task list is:: 48 49 void release_task(struct task_struct *p) 50 { 51 write_lock(&tasklist_lock); 52 list_del_rcu(&p->tasks); 53 write_unlock(&tasklist_lock); 54 call_rcu(&p->rcu, delayed_put_task_struct); 55 } 56 57When a process exits, ``release_task()`` calls ``list_del_rcu(&p->tasks)`` 58via __exit_signal() and __unhash_process() under ``tasklist_lock`` 59writer lock protection. The list_del_rcu() invocation removes 60the task from the list of all tasks. The ``tasklist_lock`` 61prevents concurrent list additions/removals from corrupting the 62list. Readers using ``for_each_process()`` are not protected with the 63``tasklist_lock``. To prevent readers from noticing changes in the list 64pointers, the ``task_struct`` object is freed only after one or more 65grace periods elapse, with the help of call_rcu(), which is invoked via 66put_task_struct_rcu_user(). This deferring of destruction ensures that 67any readers traversing the list will see valid ``p->tasks.next`` pointers 68and deletion/freeing can happen in parallel with traversal of the list. 69This pattern is also called an **existence lock**, since RCU refrains 70from invoking the delayed_put_task_struct() callback function until 71all existing readers finish, which guarantees that the ``task_struct`` 72object in question will remain in existence until after the completion 73of all RCU readers that might possibly have a reference to that object. 74 75 76Example 2: Read-Side Action Taken Outside of Lock: No In-Place Updates 77---------------------------------------------------------------------- 78 79Some reader-writer locking use cases compute a value while holding 80the read-side lock, but continue to use that value after that lock is 81released. These use cases are often good candidates for conversion 82to RCU. One prominent example involves network packet routing. 83Because the packet-routing data tracks the state of equipment outside 84of the computer, it will at times contain stale data. Therefore, once 85the route has been computed, there is no need to hold the routing table 86static during transmission of the packet. After all, you can hold the 87routing table static all you want, but that won't keep the external 88Internet from changing, and it is the state of the external Internet 89that really matters. In addition, routing entries are typically added 90or deleted, rather than being modified in place. This is a rare example 91of the finite speed of light and the non-zero size of atoms actually 92helping make synchronization be lighter weight. 93 94A straightforward example of this type of RCU use case may be found in 95the system-call auditing support. For example, a reader-writer locked 96implementation of ``audit_filter_task()`` might be as follows:: 97 98 static enum audit_state audit_filter_task(struct task_struct *tsk, char **key) 99 { 100 struct audit_entry *e; 101 enum audit_state state; 102 103 read_lock(&auditsc_lock); 104 /* Note: audit_filter_mutex held by caller. */ 105 list_for_each_entry(e, &audit_tsklist, list) { 106 if (audit_filter_rules(tsk, &e->rule, NULL, &state)) { 107 if (state == AUDIT_STATE_RECORD) 108 *key = kstrdup(e->rule.filterkey, GFP_ATOMIC); 109 read_unlock(&auditsc_lock); 110 return state; 111 } 112 } 113 read_unlock(&auditsc_lock); 114 return AUDIT_BUILD_CONTEXT; 115 } 116 117Here the list is searched under the lock, but the lock is dropped before 118the corresponding value is returned. By the time that this value is acted 119on, the list may well have been modified. This makes sense, since if 120you are turning auditing off, it is OK to audit a few extra system calls. 121 122This means that RCU can be easily applied to the read side, as follows:: 123 124 static enum audit_state audit_filter_task(struct task_struct *tsk, char **key) 125 { 126 struct audit_entry *e; 127 enum audit_state state; 128 129 rcu_read_lock(); 130 /* Note: audit_filter_mutex held by caller. */ 131 list_for_each_entry_rcu(e, &audit_tsklist, list) { 132 if (audit_filter_rules(tsk, &e->rule, NULL, &state)) { 133 if (state == AUDIT_STATE_RECORD) 134 *key = kstrdup(e->rule.filterkey, GFP_ATOMIC); 135 rcu_read_unlock(); 136 return state; 137 } 138 } 139 rcu_read_unlock(); 140 return AUDIT_BUILD_CONTEXT; 141 } 142 143The read_lock() and read_unlock() calls have become rcu_read_lock() 144and rcu_read_unlock(), respectively, and the list_for_each_entry() 145has become list_for_each_entry_rcu(). The **_rcu()** list-traversal 146primitives add READ_ONCE() and diagnostic checks for incorrect use 147outside of an RCU read-side critical section. 148 149The changes to the update side are also straightforward. A reader-writer lock 150might be used as follows for deletion and insertion in these simplified 151versions of audit_del_rule() and audit_add_rule():: 152 153 static inline int audit_del_rule(struct audit_rule *rule, 154 struct list_head *list) 155 { 156 struct audit_entry *e; 157 158 write_lock(&auditsc_lock); 159 list_for_each_entry(e, list, list) { 160 if (!audit_compare_rule(rule, &e->rule)) { 161 list_del(&e->list); 162 write_unlock(&auditsc_lock); 163 return 0; 164 } 165 } 166 write_unlock(&auditsc_lock); 167 return -EFAULT; /* No matching rule */ 168 } 169 170 static inline int audit_add_rule(struct audit_entry *entry, 171 struct list_head *list) 172 { 173 write_lock(&auditsc_lock); 174 if (entry->rule.flags & AUDIT_PREPEND) { 175 entry->rule.flags &= ~AUDIT_PREPEND; 176 list_add(&entry->list, list); 177 } else { 178 list_add_tail(&entry->list, list); 179 } 180 write_unlock(&auditsc_lock); 181 return 0; 182 } 183 184Following are the RCU equivalents for these two functions:: 185 186 static inline int audit_del_rule(struct audit_rule *rule, 187 struct list_head *list) 188 { 189 struct audit_entry *e; 190 191 /* No need to use the _rcu iterator here, since this is the only 192 * deletion routine. */ 193 list_for_each_entry(e, list, list) { 194 if (!audit_compare_rule(rule, &e->rule)) { 195 list_del_rcu(&e->list); 196 call_rcu(&e->rcu, audit_free_rule); 197 return 0; 198 } 199 } 200 return -EFAULT; /* No matching rule */ 201 } 202 203 static inline int audit_add_rule(struct audit_entry *entry, 204 struct list_head *list) 205 { 206 if (entry->rule.flags & AUDIT_PREPEND) { 207 entry->rule.flags &= ~AUDIT_PREPEND; 208 list_add_rcu(&entry->list, list); 209 } else { 210 list_add_tail_rcu(&entry->list, list); 211 } 212 return 0; 213 } 214 215Normally, the write_lock() and write_unlock() would be replaced by a 216spin_lock() and a spin_unlock(). But in this case, all callers hold 217``audit_filter_mutex``, so no additional locking is required. The 218auditsc_lock can therefore be eliminated, since use of RCU eliminates the 219need for writers to exclude readers. 220 221The list_del(), list_add(), and list_add_tail() primitives have been 222replaced by list_del_rcu(), list_add_rcu(), and list_add_tail_rcu(). 223The **_rcu()** list-manipulation primitives add memory barriers that are 224needed on weakly ordered CPUs. The list_del_rcu() primitive omits the 225pointer poisoning debug-assist code that would otherwise cause concurrent 226readers to fail spectacularly. 227 228So, when readers can tolerate stale data and when entries are either added or 229deleted, without in-place modification, it is very easy to use RCU! 230 231 232Example 3: Handling In-Place Updates 233------------------------------------ 234 235The system-call auditing code does not update auditing rules in place. However, 236if it did, the reader-writer-locked code to do so might look as follows 237(assuming only ``field_count`` is updated, otherwise, the added fields would 238need to be filled in):: 239 240 static inline int audit_upd_rule(struct audit_rule *rule, 241 struct list_head *list, 242 __u32 newaction, 243 __u32 newfield_count) 244 { 245 struct audit_entry *e; 246 struct audit_entry *ne; 247 248 write_lock(&auditsc_lock); 249 /* Note: audit_filter_mutex held by caller. */ 250 list_for_each_entry(e, list, list) { 251 if (!audit_compare_rule(rule, &e->rule)) { 252 e->rule.action = newaction; 253 e->rule.field_count = newfield_count; 254 write_unlock(&auditsc_lock); 255 return 0; 256 } 257 } 258 write_unlock(&auditsc_lock); 259 return -EFAULT; /* No matching rule */ 260 } 261 262The RCU version creates a copy, updates the copy, then replaces the old 263entry with the newly updated entry. This sequence of actions, allowing 264concurrent reads while making a copy to perform an update, is what gives 265RCU (*read-copy update*) its name. 266 267The RCU version of audit_upd_rule() is as follows:: 268 269 static inline int audit_upd_rule(struct audit_rule *rule, 270 struct list_head *list, 271 __u32 newaction, 272 __u32 newfield_count) 273 { 274 struct audit_entry *e; 275 struct audit_entry *ne; 276 277 list_for_each_entry(e, list, list) { 278 if (!audit_compare_rule(rule, &e->rule)) { 279 ne = kmalloc(sizeof(*entry), GFP_ATOMIC); 280 if (ne == NULL) 281 return -ENOMEM; 282 audit_copy_rule(&ne->rule, &e->rule); 283 ne->rule.action = newaction; 284 ne->rule.field_count = newfield_count; 285 list_replace_rcu(&e->list, &ne->list); 286 call_rcu(&e->rcu, audit_free_rule); 287 return 0; 288 } 289 } 290 return -EFAULT; /* No matching rule */ 291 } 292 293Again, this assumes that the caller holds ``audit_filter_mutex``. Normally, the 294writer lock would become a spinlock in this sort of code. 295 296The update_lsm_rule() does something very similar, for those who would 297prefer to look at real Linux-kernel code. 298 299Another use of this pattern can be found in the openswitch driver's *connection 300tracking table* code in ``ct_limit_set()``. The table holds connection tracking 301entries and has a limit on the maximum entries. There is one such table 302per-zone and hence one *limit* per zone. The zones are mapped to their limits 303through a hashtable using an RCU-managed hlist for the hash chains. When a new 304limit is set, a new limit object is allocated and ``ct_limit_set()`` is called 305to replace the old limit object with the new one using list_replace_rcu(). 306The old limit object is then freed after a grace period using kfree_rcu(). 307 308 309Example 4: Eliminating Stale Data 310--------------------------------- 311 312The auditing example above tolerates stale data, as do most algorithms 313that are tracking external state. After all, given there is a delay 314from the time the external state changes before Linux becomes aware 315of the change, and so as noted earlier, a small quantity of additional 316RCU-induced staleness is generally not a problem. 317 318However, there are many examples where stale data cannot be tolerated. 319One example in the Linux kernel is the System V IPC (see the shm_lock() 320function in ipc/shm.c). This code checks a *deleted* flag under a 321per-entry spinlock, and, if the *deleted* flag is set, pretends that the 322entry does not exist. For this to be helpful, the search function must 323return holding the per-entry spinlock, as shm_lock() does in fact do. 324 325.. _quick_quiz: 326 327Quick Quiz: 328 For the deleted-flag technique to be helpful, why is it necessary 329 to hold the per-entry lock while returning from the search function? 330 331:ref:`Answer to Quick Quiz <quick_quiz_answer>` 332 333If the system-call audit module were to ever need to reject stale data, one way 334to accomplish this would be to add a ``deleted`` flag and a ``lock`` spinlock to the 335``audit_entry`` structure, and modify audit_filter_task() as follows:: 336 337 static enum audit_state audit_filter_task(struct task_struct *tsk) 338 { 339 struct audit_entry *e; 340 enum audit_state state; 341 342 rcu_read_lock(); 343 list_for_each_entry_rcu(e, &audit_tsklist, list) { 344 if (audit_filter_rules(tsk, &e->rule, NULL, &state)) { 345 spin_lock(&e->lock); 346 if (e->deleted) { 347 spin_unlock(&e->lock); 348 rcu_read_unlock(); 349 return AUDIT_BUILD_CONTEXT; 350 } 351 rcu_read_unlock(); 352 if (state == AUDIT_STATE_RECORD) 353 *key = kstrdup(e->rule.filterkey, GFP_ATOMIC); 354 return state; 355 } 356 } 357 rcu_read_unlock(); 358 return AUDIT_BUILD_CONTEXT; 359 } 360 361The ``audit_del_rule()`` function would need to set the ``deleted`` flag under the 362spinlock as follows:: 363 364 static inline int audit_del_rule(struct audit_rule *rule, 365 struct list_head *list) 366 { 367 struct audit_entry *e; 368 369 /* No need to use the _rcu iterator here, since this 370 * is the only deletion routine. */ 371 list_for_each_entry(e, list, list) { 372 if (!audit_compare_rule(rule, &e->rule)) { 373 spin_lock(&e->lock); 374 list_del_rcu(&e->list); 375 e->deleted = 1; 376 spin_unlock(&e->lock); 377 call_rcu(&e->rcu, audit_free_rule); 378 return 0; 379 } 380 } 381 return -EFAULT; /* No matching rule */ 382 } 383 384This too assumes that the caller holds ``audit_filter_mutex``. 385 386Note that this example assumes that entries are only added and deleted. 387Additional mechanism is required to deal correctly with the update-in-place 388performed by audit_upd_rule(). For one thing, audit_upd_rule() would 389need to hold the locks of both the old ``audit_entry`` and its replacement 390while executing the list_replace_rcu(). 391 392 393Example 5: Skipping Stale Objects 394--------------------------------- 395 396For some use cases, reader performance can be improved by skipping 397stale objects during read-side list traversal, where stale objects 398are those that will be removed and destroyed after one or more grace 399periods. One such example can be found in the timerfd subsystem. When a 400``CLOCK_REALTIME`` clock is reprogrammed (for example due to setting 401of the system time) then all programmed ``timerfds`` that depend on 402this clock get triggered and processes waiting on them are awakened in 403advance of their scheduled expiry. To facilitate this, all such timers 404are added to an RCU-managed ``cancel_list`` when they are setup in 405``timerfd_setup_cancel()``:: 406 407 static void timerfd_setup_cancel(struct timerfd_ctx *ctx, int flags) 408 { 409 spin_lock(&ctx->cancel_lock); 410 if ((ctx->clockid == CLOCK_REALTIME || 411 ctx->clockid == CLOCK_REALTIME_ALARM) && 412 (flags & TFD_TIMER_ABSTIME) && (flags & TFD_TIMER_CANCEL_ON_SET)) { 413 if (!ctx->might_cancel) { 414 ctx->might_cancel = true; 415 spin_lock(&cancel_lock); 416 list_add_rcu(&ctx->clist, &cancel_list); 417 spin_unlock(&cancel_lock); 418 } 419 } else { 420 __timerfd_remove_cancel(ctx); 421 } 422 spin_unlock(&ctx->cancel_lock); 423 } 424 425When a timerfd is freed (fd is closed), then the ``might_cancel`` 426flag of the timerfd object is cleared, the object removed from the 427``cancel_list`` and destroyed, as shown in this simplified and inlined 428version of timerfd_release():: 429 430 int timerfd_release(struct inode *inode, struct file *file) 431 { 432 struct timerfd_ctx *ctx = file->private_data; 433 434 spin_lock(&ctx->cancel_lock); 435 if (ctx->might_cancel) { 436 ctx->might_cancel = false; 437 spin_lock(&cancel_lock); 438 list_del_rcu(&ctx->clist); 439 spin_unlock(&cancel_lock); 440 } 441 spin_unlock(&ctx->cancel_lock); 442 443 if (isalarm(ctx)) 444 alarm_cancel(&ctx->t.alarm); 445 else 446 hrtimer_cancel(&ctx->t.tmr); 447 kfree_rcu(ctx, rcu); 448 return 0; 449 } 450 451If the ``CLOCK_REALTIME`` clock is set, for example by a time server, the 452hrtimer framework calls ``timerfd_clock_was_set()`` which walks the 453``cancel_list`` and wakes up processes waiting on the timerfd. While iterating 454the ``cancel_list``, the ``might_cancel`` flag is consulted to skip stale 455objects:: 456 457 void timerfd_clock_was_set(void) 458 { 459 ktime_t moffs = ktime_mono_to_real(0); 460 struct timerfd_ctx *ctx; 461 unsigned long flags; 462 463 rcu_read_lock(); 464 list_for_each_entry_rcu(ctx, &cancel_list, clist) { 465 if (!ctx->might_cancel) 466 continue; 467 spin_lock_irqsave(&ctx->wqh.lock, flags); 468 if (ctx->moffs != moffs) { 469 ctx->moffs = KTIME_MAX; 470 ctx->ticks++; 471 wake_up_locked_poll(&ctx->wqh, EPOLLIN); 472 } 473 spin_unlock_irqrestore(&ctx->wqh.lock, flags); 474 } 475 rcu_read_unlock(); 476 } 477 478The key point is that because RCU-protected traversal of the 479``cancel_list`` happens concurrently with object addition and removal, 480sometimes the traversal can access an object that has been removed from 481the list. In this example, a flag is used to skip such objects. 482 483 484Summary 485------- 486 487Read-mostly list-based data structures that can tolerate stale data are 488the most amenable to use of RCU. The simplest case is where entries are 489either added or deleted from the data structure (or atomically modified 490in place), but non-atomic in-place modifications can be handled by making 491a copy, updating the copy, then replacing the original with the copy. 492If stale data cannot be tolerated, then a *deleted* flag may be used 493in conjunction with a per-entry spinlock in order to allow the search 494function to reject newly deleted data. 495 496.. _quick_quiz_answer: 497 498Answer to Quick Quiz: 499 For the deleted-flag technique to be helpful, why is it necessary 500 to hold the per-entry lock while returning from the search function? 501 502 If the search function drops the per-entry lock before returning, 503 then the caller will be processing stale data in any case. If it 504 is really OK to be processing stale data, then you don't need a 505 *deleted* flag. If processing stale data really is a problem, 506 then you need to hold the per-entry lock across all of the code 507 that uses the value that was returned. 508 509:ref:`Back to Quick Quiz <quick_quiz>` 510