Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | Using RCU to Protect Read-Mostly Linked Lists |
| 2 | |
| 3 | |
| 4 | One of the best applications of RCU is to protect read-mostly linked lists |
| 5 | ("struct list_head" in list.h). One big advantage of this approach |
| 6 | is that all of the required memory barriers are included for you in |
| 7 | the list macros. This document describes several applications of RCU, |
| 8 | with the best fits first. |
| 9 | |
| 10 | |
| 11 | Example 1: Read-Side Action Taken Outside of Lock, No In-Place Updates |
| 12 | |
| 13 | The best applications are cases where, if reader-writer locking were |
| 14 | used, the read-side lock would be dropped before taking any action |
| 15 | based on the results of the search. The most celebrated example is |
| 16 | the routing table. Because the routing table is tracking the state of |
| 17 | equipment outside of the computer, it will at times contain stale data. |
| 18 | Therefore, once the route has been computed, there is no need to hold |
| 19 | the routing table static during transmission of the packet. After all, |
| 20 | you can hold the routing table static all you want, but that won't keep |
| 21 | the external Internet from changing, and it is the state of the external |
| 22 | Internet that really matters. In addition, routing entries are typically |
| 23 | added or deleted, rather than being modified in place. |
| 24 | |
| 25 | A straightforward example of this use of RCU may be found in the |
| 26 | system-call auditing support. For example, a reader-writer locked |
| 27 | implementation of audit_filter_task() might be as follows: |
| 28 | |
| 29 | static enum audit_state audit_filter_task(struct task_struct *tsk) |
| 30 | { |
| 31 | struct audit_entry *e; |
| 32 | enum audit_state state; |
| 33 | |
| 34 | read_lock(&auditsc_lock); |
Paul E. McKenney | a83f1fe | 2005-05-01 08:59:05 -0700 | [diff] [blame] | 35 | /* Note: audit_netlink_sem held by caller. */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 36 | list_for_each_entry(e, &audit_tsklist, list) { |
| 37 | if (audit_filter_rules(tsk, &e->rule, NULL, &state)) { |
| 38 | read_unlock(&auditsc_lock); |
| 39 | return state; |
| 40 | } |
| 41 | } |
| 42 | read_unlock(&auditsc_lock); |
| 43 | return AUDIT_BUILD_CONTEXT; |
| 44 | } |
| 45 | |
| 46 | Here the list is searched under the lock, but the lock is dropped before |
| 47 | the corresponding value is returned. By the time that this value is acted |
| 48 | on, the list may well have been modified. This makes sense, since if |
| 49 | you are turning auditing off, it is OK to audit a few extra system calls. |
| 50 | |
| 51 | This means that RCU can be easily applied to the read side, as follows: |
| 52 | |
| 53 | static enum audit_state audit_filter_task(struct task_struct *tsk) |
| 54 | { |
| 55 | struct audit_entry *e; |
| 56 | enum audit_state state; |
| 57 | |
| 58 | rcu_read_lock(); |
Paul E. McKenney | a83f1fe | 2005-05-01 08:59:05 -0700 | [diff] [blame] | 59 | /* Note: audit_netlink_sem held by caller. */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 60 | list_for_each_entry_rcu(e, &audit_tsklist, list) { |
| 61 | if (audit_filter_rules(tsk, &e->rule, NULL, &state)) { |
| 62 | rcu_read_unlock(); |
| 63 | return state; |
| 64 | } |
| 65 | } |
| 66 | rcu_read_unlock(); |
| 67 | return AUDIT_BUILD_CONTEXT; |
| 68 | } |
| 69 | |
| 70 | The read_lock() and read_unlock() calls have become rcu_read_lock() |
| 71 | and rcu_read_unlock(), respectively, and the list_for_each_entry() has |
| 72 | become list_for_each_entry_rcu(). The _rcu() list-traversal primitives |
| 73 | insert the read-side memory barriers that are required on DEC Alpha CPUs. |
| 74 | |
| 75 | The changes to the update side are also straightforward. A reader-writer |
| 76 | lock might be used as follows for deletion and insertion: |
| 77 | |
| 78 | static inline int audit_del_rule(struct audit_rule *rule, |
| 79 | struct list_head *list) |
| 80 | { |
| 81 | struct audit_entry *e; |
| 82 | |
| 83 | write_lock(&auditsc_lock); |
| 84 | list_for_each_entry(e, list, list) { |
| 85 | if (!audit_compare_rule(rule, &e->rule)) { |
| 86 | list_del(&e->list); |
| 87 | write_unlock(&auditsc_lock); |
| 88 | return 0; |
| 89 | } |
| 90 | } |
| 91 | write_unlock(&auditsc_lock); |
| 92 | return -EFAULT; /* No matching rule */ |
| 93 | } |
| 94 | |
| 95 | static inline int audit_add_rule(struct audit_entry *entry, |
| 96 | struct list_head *list) |
| 97 | { |
| 98 | write_lock(&auditsc_lock); |
| 99 | if (entry->rule.flags & AUDIT_PREPEND) { |
| 100 | entry->rule.flags &= ~AUDIT_PREPEND; |
| 101 | list_add(&entry->list, list); |
| 102 | } else { |
| 103 | list_add_tail(&entry->list, list); |
| 104 | } |
| 105 | write_unlock(&auditsc_lock); |
| 106 | return 0; |
| 107 | } |
| 108 | |
| 109 | Following are the RCU equivalents for these two functions: |
| 110 | |
| 111 | static inline int audit_del_rule(struct audit_rule *rule, |
| 112 | struct list_head *list) |
| 113 | { |
| 114 | struct audit_entry *e; |
| 115 | |
| 116 | /* Do not use the _rcu iterator here, since this is the only |
| 117 | * deletion routine. */ |
| 118 | list_for_each_entry(e, list, list) { |
| 119 | if (!audit_compare_rule(rule, &e->rule)) { |
| 120 | list_del_rcu(&e->list); |
| 121 | call_rcu(&e->rcu, audit_free_rule, e); |
| 122 | return 0; |
| 123 | } |
| 124 | } |
| 125 | return -EFAULT; /* No matching rule */ |
| 126 | } |
| 127 | |
| 128 | static inline int audit_add_rule(struct audit_entry *entry, |
| 129 | struct list_head *list) |
| 130 | { |
| 131 | if (entry->rule.flags & AUDIT_PREPEND) { |
| 132 | entry->rule.flags &= ~AUDIT_PREPEND; |
| 133 | list_add_rcu(&entry->list, list); |
| 134 | } else { |
| 135 | list_add_tail_rcu(&entry->list, list); |
| 136 | } |
| 137 | return 0; |
| 138 | } |
| 139 | |
| 140 | Normally, the write_lock() and write_unlock() would be replaced by |
| 141 | a spin_lock() and a spin_unlock(), but in this case, all callers hold |
| 142 | audit_netlink_sem, so no additional locking is required. The auditsc_lock |
| 143 | can therefore be eliminated, since use of RCU eliminates the need for |
Paul E. McKenney | a83f1fe | 2005-05-01 08:59:05 -0700 | [diff] [blame] | 144 | writers to exclude readers. Normally, the write_lock() calls would |
| 145 | be converted into spin_lock() calls. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 146 | |
| 147 | The list_del(), list_add(), and list_add_tail() primitives have been |
| 148 | replaced by list_del_rcu(), list_add_rcu(), and list_add_tail_rcu(). |
| 149 | The _rcu() list-manipulation primitives add memory barriers that are |
Paul E. McKenney | a83f1fe | 2005-05-01 08:59:05 -0700 | [diff] [blame] | 150 | needed on weakly ordered CPUs (most of them!). The list_del_rcu() |
| 151 | primitive omits the pointer poisoning debug-assist code that would |
| 152 | otherwise cause concurrent readers to fail spectacularly. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 153 | |
| 154 | So, when readers can tolerate stale data and when entries are either added |
| 155 | or deleted, without in-place modification, it is very easy to use RCU! |
| 156 | |
| 157 | |
| 158 | Example 2: Handling In-Place Updates |
| 159 | |
| 160 | The system-call auditing code does not update auditing rules in place. |
| 161 | However, if it did, reader-writer-locked code to do so might look as |
| 162 | follows (presumably, the field_count is only permitted to decrease, |
| 163 | otherwise, the added fields would need to be filled in): |
| 164 | |
| 165 | static inline int audit_upd_rule(struct audit_rule *rule, |
| 166 | struct list_head *list, |
| 167 | __u32 newaction, |
| 168 | __u32 newfield_count) |
| 169 | { |
| 170 | struct audit_entry *e; |
| 171 | struct audit_newentry *ne; |
| 172 | |
| 173 | write_lock(&auditsc_lock); |
Paul E. McKenney | a83f1fe | 2005-05-01 08:59:05 -0700 | [diff] [blame] | 174 | /* Note: audit_netlink_sem held by caller. */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 175 | list_for_each_entry(e, list, list) { |
| 176 | if (!audit_compare_rule(rule, &e->rule)) { |
| 177 | e->rule.action = newaction; |
| 178 | e->rule.file_count = newfield_count; |
| 179 | write_unlock(&auditsc_lock); |
| 180 | return 0; |
| 181 | } |
| 182 | } |
| 183 | write_unlock(&auditsc_lock); |
| 184 | return -EFAULT; /* No matching rule */ |
| 185 | } |
| 186 | |
| 187 | The RCU version creates a copy, updates the copy, then replaces the old |
| 188 | entry with the newly updated entry. This sequence of actions, allowing |
| 189 | concurrent reads while doing a copy to perform an update, is what gives |
| 190 | RCU ("read-copy update") its name. The RCU code is as follows: |
| 191 | |
| 192 | static inline int audit_upd_rule(struct audit_rule *rule, |
| 193 | struct list_head *list, |
| 194 | __u32 newaction, |
| 195 | __u32 newfield_count) |
| 196 | { |
| 197 | struct audit_entry *e; |
| 198 | struct audit_newentry *ne; |
| 199 | |
| 200 | list_for_each_entry(e, list, list) { |
| 201 | if (!audit_compare_rule(rule, &e->rule)) { |
| 202 | ne = kmalloc(sizeof(*entry), GFP_ATOMIC); |
| 203 | if (ne == NULL) |
| 204 | return -ENOMEM; |
| 205 | audit_copy_rule(&ne->rule, &e->rule); |
| 206 | ne->rule.action = newaction; |
| 207 | ne->rule.file_count = newfield_count; |
Paul E. McKenney | a83f1fe | 2005-05-01 08:59:05 -0700 | [diff] [blame] | 208 | list_replace_rcu(e, ne); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 209 | call_rcu(&e->rcu, audit_free_rule, e); |
| 210 | return 0; |
| 211 | } |
| 212 | } |
| 213 | return -EFAULT; /* No matching rule */ |
| 214 | } |
| 215 | |
| 216 | Again, this assumes that the caller holds audit_netlink_sem. Normally, |
| 217 | the reader-writer lock would become a spinlock in this sort of code. |
| 218 | |
| 219 | |
| 220 | Example 3: Eliminating Stale Data |
| 221 | |
| 222 | The auditing examples above tolerate stale data, as do most algorithms |
| 223 | that are tracking external state. Because there is a delay from the |
| 224 | time the external state changes before Linux becomes aware of the change, |
| 225 | additional RCU-induced staleness is normally not a problem. |
| 226 | |
| 227 | However, there are many examples where stale data cannot be tolerated. |
| 228 | One example in the Linux kernel is the System V IPC (see the ipc_lock() |
| 229 | function in ipc/util.c). This code checks a "deleted" flag under a |
| 230 | per-entry spinlock, and, if the "deleted" flag is set, pretends that the |
| 231 | entry does not exist. For this to be helpful, the search function must |
| 232 | return holding the per-entry spinlock, as ipc_lock() does in fact do. |
| 233 | |
| 234 | Quick Quiz: Why does the search function need to return holding the |
Paul E. McKenney | d19720a | 2006-02-01 03:06:42 -0800 | [diff] [blame] | 235 | per-entry lock for this deleted-flag technique to be helpful? |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 236 | |
| 237 | If the system-call audit module were to ever need to reject stale data, |
| 238 | one way to accomplish this would be to add a "deleted" flag and a "lock" |
| 239 | spinlock to the audit_entry structure, and modify audit_filter_task() |
| 240 | as follows: |
| 241 | |
| 242 | static enum audit_state audit_filter_task(struct task_struct *tsk) |
| 243 | { |
| 244 | struct audit_entry *e; |
| 245 | enum audit_state state; |
| 246 | |
| 247 | rcu_read_lock(); |
| 248 | list_for_each_entry_rcu(e, &audit_tsklist, list) { |
| 249 | if (audit_filter_rules(tsk, &e->rule, NULL, &state)) { |
| 250 | spin_lock(&e->lock); |
| 251 | if (e->deleted) { |
| 252 | spin_unlock(&e->lock); |
| 253 | rcu_read_unlock(); |
| 254 | return AUDIT_BUILD_CONTEXT; |
| 255 | } |
| 256 | rcu_read_unlock(); |
| 257 | return state; |
| 258 | } |
| 259 | } |
| 260 | rcu_read_unlock(); |
| 261 | return AUDIT_BUILD_CONTEXT; |
| 262 | } |
| 263 | |
| 264 | Note that this example assumes that entries are only added and deleted. |
| 265 | Additional mechanism is required to deal correctly with the |
| 266 | update-in-place performed by audit_upd_rule(). For one thing, |
| 267 | audit_upd_rule() would need additional memory barriers to ensure |
| 268 | that the list_add_rcu() was really executed before the list_del_rcu(). |
| 269 | |
| 270 | The audit_del_rule() function would need to set the "deleted" |
| 271 | flag under the spinlock as follows: |
| 272 | |
| 273 | static inline int audit_del_rule(struct audit_rule *rule, |
| 274 | struct list_head *list) |
| 275 | { |
| 276 | struct audit_entry *e; |
| 277 | |
Paul E. McKenney | d19720a | 2006-02-01 03:06:42 -0800 | [diff] [blame] | 278 | /* Do not need to use the _rcu iterator here, since this |
| 279 | * is the only deletion routine. */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 280 | list_for_each_entry(e, list, list) { |
| 281 | if (!audit_compare_rule(rule, &e->rule)) { |
| 282 | spin_lock(&e->lock); |
| 283 | list_del_rcu(&e->list); |
| 284 | e->deleted = 1; |
| 285 | spin_unlock(&e->lock); |
| 286 | call_rcu(&e->rcu, audit_free_rule, e); |
| 287 | return 0; |
| 288 | } |
| 289 | } |
| 290 | return -EFAULT; /* No matching rule */ |
| 291 | } |
| 292 | |
| 293 | |
| 294 | Summary |
| 295 | |
| 296 | Read-mostly list-based data structures that can tolerate stale data are |
| 297 | the most amenable to use of RCU. The simplest case is where entries are |
| 298 | either added or deleted from the data structure (or atomically modified |
| 299 | in place), but non-atomic in-place modifications can be handled by making |
| 300 | a copy, updating the copy, then replacing the original with the copy. |
| 301 | If stale data cannot be tolerated, then a "deleted" flag may be used |
| 302 | in conjunction with a per-entry spinlock in order to allow the search |
| 303 | function to reject newly deleted data. |
| 304 | |
| 305 | |
| 306 | Answer to Quick Quiz |
Paul E. McKenney | d19720a | 2006-02-01 03:06:42 -0800 | [diff] [blame] | 307 | Why does the search function need to return holding the per-entry |
| 308 | lock for this deleted-flag technique to be helpful? |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 309 | |
Paul E. McKenney | d19720a | 2006-02-01 03:06:42 -0800 | [diff] [blame] | 310 | If the search function drops the per-entry lock before returning, |
| 311 | then the caller will be processing stale data in any case. If it |
| 312 | is really OK to be processing stale data, then you don't need a |
| 313 | "deleted" flag. If processing stale data really is a problem, |
| 314 | then you need to hold the per-entry lock across all of the code |
| 315 | that uses the value that was returned. |