PageRenderTime 10ms CodeModel.GetById 1ms app.highlight 5ms RepoModel.GetById 1ms app.codeStats 0ms

/Documentation/RCU/listRCU.txt

https://bitbucket.org/evzijst/gittest
Plain Text | 307 lines | 256 code | 51 blank | 0 comment | 0 complexity | 7ba3fb72cabebc2c42a63bacb844bc39 MD5 | raw file
  1Using RCU to Protect Read-Mostly Linked Lists
  2
  3
  4One 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
  6is that all of the required memory barriers are included for you in
  7the list macros.  This document describes several applications of RCU,
  8with the best fits first.
  9
 10
 11Example 1: Read-Side Action Taken Outside of Lock, No In-Place Updates
 12
 13The best applications are cases where, if reader-writer locking were
 14used, the read-side lock would be dropped before taking any action
 15based on the results of the search.  The most celebrated example is
 16the routing table.  Because the routing table is tracking the state of
 17equipment outside of the computer, it will at times contain stale data.
 18Therefore, once the route has been computed, there is no need to hold
 19the routing table static during transmission of the packet.  After all,
 20you can hold the routing table static all you want, but that won't keep
 21the external Internet from changing, and it is the state of the external
 22Internet that really matters.  In addition, routing entries are typically
 23added or deleted, rather than being modified in place.
 24
 25A straightforward example of this use of RCU may be found in the
 26system-call auditing support.  For example, a reader-writer locked
 27implementation 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);
 35		list_for_each_entry(e, &audit_tsklist, list) {
 36			if (audit_filter_rules(tsk, &e->rule, NULL, &state)) {
 37				read_unlock(&auditsc_lock);
 38				return state;
 39			}
 40		}
 41		read_unlock(&auditsc_lock);
 42		return AUDIT_BUILD_CONTEXT;
 43	}
 44
 45Here the list is searched under the lock, but the lock is dropped before
 46the corresponding value is returned.  By the time that this value is acted
 47on, the list may well have been modified.  This makes sense, since if
 48you are turning auditing off, it is OK to audit a few extra system calls.
 49
 50This means that RCU can be easily applied to the read side, as follows:
 51
 52	static enum audit_state audit_filter_task(struct task_struct *tsk)
 53	{
 54		struct audit_entry *e;
 55		enum audit_state   state;
 56
 57		rcu_read_lock();
 58		list_for_each_entry_rcu(e, &audit_tsklist, list) {
 59			if (audit_filter_rules(tsk, &e->rule, NULL, &state)) {
 60				rcu_read_unlock();
 61				return state;
 62			}
 63		}
 64		rcu_read_unlock();
 65		return AUDIT_BUILD_CONTEXT;
 66	}
 67
 68The read_lock() and read_unlock() calls have become rcu_read_lock()
 69and rcu_read_unlock(), respectively, and the list_for_each_entry() has
 70become list_for_each_entry_rcu().  The _rcu() list-traversal primitives
 71insert the read-side memory barriers that are required on DEC Alpha CPUs.
 72
 73The changes to the update side are also straightforward.  A reader-writer
 74lock might be used as follows for deletion and insertion:
 75
 76	static inline int audit_del_rule(struct audit_rule *rule,
 77					 struct list_head *list)
 78	{
 79		struct audit_entry  *e;
 80
 81		write_lock(&auditsc_lock);
 82		list_for_each_entry(e, list, list) {
 83			if (!audit_compare_rule(rule, &e->rule)) {
 84				list_del(&e->list);
 85				write_unlock(&auditsc_lock);
 86				return 0;
 87			}
 88		}
 89		write_unlock(&auditsc_lock);
 90		return -EFAULT;		/* No matching rule */
 91	}
 92
 93	static inline int audit_add_rule(struct audit_entry *entry,
 94					 struct list_head *list)
 95	{
 96		write_lock(&auditsc_lock);
 97		if (entry->rule.flags & AUDIT_PREPEND) {
 98			entry->rule.flags &= ~AUDIT_PREPEND;
 99			list_add(&entry->list, list);
100		} else {
101			list_add_tail(&entry->list, list);
102		}
103		write_unlock(&auditsc_lock);
104		return 0;
105	}
106
107Following are the RCU equivalents for these two functions:
108
109	static inline int audit_del_rule(struct audit_rule *rule,
110					 struct list_head *list)
111	{
112		struct audit_entry  *e;
113
114		/* Do not use the _rcu iterator here, since this is the only
115		 * deletion routine. */
116		list_for_each_entry(e, list, list) {
117			if (!audit_compare_rule(rule, &e->rule)) {
118				list_del_rcu(&e->list);
119				call_rcu(&e->rcu, audit_free_rule, e);
120				return 0;
121			}
122		}
123		return -EFAULT;		/* No matching rule */
124	}
125
126	static inline int audit_add_rule(struct audit_entry *entry,
127					 struct list_head *list)
128	{
129		if (entry->rule.flags & AUDIT_PREPEND) {
130			entry->rule.flags &= ~AUDIT_PREPEND;
131			list_add_rcu(&entry->list, list);
132		} else {
133			list_add_tail_rcu(&entry->list, list);
134		}
135		return 0;
136	}
137
138Normally, the write_lock() and write_unlock() would be replaced by
139a spin_lock() and a spin_unlock(), but in this case, all callers hold
140audit_netlink_sem, so no additional locking is required.  The auditsc_lock
141can therefore be eliminated, since use of RCU eliminates the need for
142writers to exclude readers.
143
144The list_del(), list_add(), and list_add_tail() primitives have been
145replaced by list_del_rcu(), list_add_rcu(), and list_add_tail_rcu().
146The _rcu() list-manipulation primitives add memory barriers that are
147needed on weakly ordered CPUs (most of them!).
148
149So, when readers can tolerate stale data and when entries are either added
150or deleted, without in-place modification, it is very easy to use RCU!
151
152
153Example 2: Handling In-Place Updates
154
155The system-call auditing code does not update auditing rules in place.
156However, if it did, reader-writer-locked code to do so might look as
157follows (presumably, the field_count is only permitted to decrease,
158otherwise, the added fields would need to be filled in):
159
160	static inline int audit_upd_rule(struct audit_rule *rule,
161					 struct list_head *list,
162					 __u32 newaction,
163					 __u32 newfield_count)
164	{
165		struct audit_entry  *e;
166		struct audit_newentry *ne;
167
168		write_lock(&auditsc_lock);
169		list_for_each_entry(e, list, list) {
170			if (!audit_compare_rule(rule, &e->rule)) {
171				e->rule.action = newaction;
172				e->rule.file_count = newfield_count;
173				write_unlock(&auditsc_lock);
174				return 0;
175			}
176		}
177		write_unlock(&auditsc_lock);
178		return -EFAULT;		/* No matching rule */
179	}
180
181The RCU version creates a copy, updates the copy, then replaces the old
182entry with the newly updated entry.  This sequence of actions, allowing
183concurrent reads while doing a copy to perform an update, is what gives
184RCU ("read-copy update") its name.  The RCU code is as follows:
185
186	static inline int audit_upd_rule(struct audit_rule *rule,
187					 struct list_head *list,
188					 __u32 newaction,
189					 __u32 newfield_count)
190	{
191		struct audit_entry  *e;
192		struct audit_newentry *ne;
193
194		list_for_each_entry(e, list, list) {
195			if (!audit_compare_rule(rule, &e->rule)) {
196				ne = kmalloc(sizeof(*entry), GFP_ATOMIC);
197				if (ne == NULL)
198					return -ENOMEM;
199				audit_copy_rule(&ne->rule, &e->rule);
200				ne->rule.action = newaction;
201				ne->rule.file_count = newfield_count;
202				list_add_rcu(ne, e);
203				list_del(e);
204				call_rcu(&e->rcu, audit_free_rule, e);
205				return 0;
206			}
207		}
208		return -EFAULT;		/* No matching rule */
209	}
210
211Again, this assumes that the caller holds audit_netlink_sem.  Normally,
212the reader-writer lock would become a spinlock in this sort of code.
213
214
215Example 3: Eliminating Stale Data
216
217The auditing examples above tolerate stale data, as do most algorithms
218that are tracking external state.  Because there is a delay from the
219time the external state changes before Linux becomes aware of the change,
220additional RCU-induced staleness is normally not a problem.
221
222However, there are many examples where stale data cannot be tolerated.
223One example in the Linux kernel is the System V IPC (see the ipc_lock()
224function in ipc/util.c).  This code checks a "deleted" flag under a
225per-entry spinlock, and, if the "deleted" flag is set, pretends that the
226entry does not exist.  For this to be helpful, the search function must
227return holding the per-entry spinlock, as ipc_lock() does in fact do.
228
229Quick Quiz:  Why does the search function need to return holding the
230per-entry lock for this deleted-flag technique to be helpful?
231
232If the system-call audit module were to ever need to reject stale data,
233one way to accomplish this would be to add a "deleted" flag and a "lock"
234spinlock to the audit_entry structure, and modify audit_filter_task()
235as follows:
236
237	static enum audit_state audit_filter_task(struct task_struct *tsk)
238	{
239		struct audit_entry *e;
240		enum audit_state   state;
241
242		rcu_read_lock();
243		list_for_each_entry_rcu(e, &audit_tsklist, list) {
244			if (audit_filter_rules(tsk, &e->rule, NULL, &state)) {
245				spin_lock(&e->lock);
246				if (e->deleted) {
247					spin_unlock(&e->lock);
248					rcu_read_unlock();
249					return AUDIT_BUILD_CONTEXT;
250				}
251				rcu_read_unlock();
252				return state;
253			}
254		}
255		rcu_read_unlock();
256		return AUDIT_BUILD_CONTEXT;
257	}
258
259Note that this example assumes that entries are only added and deleted.
260Additional mechanism is required to deal correctly with the
261update-in-place performed by audit_upd_rule().  For one thing,
262audit_upd_rule() would need additional memory barriers to ensure
263that the list_add_rcu() was really executed before the list_del_rcu().
264
265The audit_del_rule() function would need to set the "deleted"
266flag under the spinlock as follows:
267
268	static inline int audit_del_rule(struct audit_rule *rule,
269					 struct list_head *list)
270	{
271		struct audit_entry  *e;
272
273		/* Do not use the _rcu iterator here, since this is the only
274		 * deletion routine. */
275		list_for_each_entry(e, list, list) {
276			if (!audit_compare_rule(rule, &e->rule)) {
277				spin_lock(&e->lock);
278				list_del_rcu(&e->list);
279				e->deleted = 1;
280				spin_unlock(&e->lock);
281				call_rcu(&e->rcu, audit_free_rule, e);
282				return 0;
283			}
284		}
285		return -EFAULT;		/* No matching rule */
286	}
287
288
289Summary
290
291Read-mostly list-based data structures that can tolerate stale data are
292the most amenable to use of RCU.  The simplest case is where entries are
293either added or deleted from the data structure (or atomically modified
294in place), but non-atomic in-place modifications can be handled by making
295a copy, updating the copy, then replacing the original with the copy.
296If stale data cannot be tolerated, then a "deleted" flag may be used
297in conjunction with a per-entry spinlock in order to allow the search
298function to reject newly deleted data.
299
300
301Answer to Quick Quiz
302
303If the search function drops the per-entry lock before returning, then
304the caller will be processing stale data in any case.  If it is really
305OK to be processing stale data, then you don't need a "deleted" flag.
306If processing stale data really is a problem, then you need to hold the
307per-entry lock across all of the code that uses the value looked up.