blob: ec407d7663980f280f841945e355551e5f784d4e [file] [log] [blame]
Glenn Kasten98afa532013-04-15 14:02:36 -07001page.title=Avoiding Priority Inversion
2@jd:body
3
Clay Murphybc92aea2014-10-16 10:13:18 -07004<!--
5 Copyright 2013 The Android Open Source Project
6
7 Licensed under the Apache License, Version 2.0 (the "License");
8 you may not use this file except in compliance with the License.
9 You may obtain a copy of the License at
10
11 http://www.apache.org/licenses/LICENSE-2.0
12
13 Unless required by applicable law or agreed to in writing, software
14 distributed under the License is distributed on an "AS IS" BASIS,
15 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
16 See the License for the specific language governing permissions and
17 limitations under the License.
18-->
Glenn Kasten98afa532013-04-15 14:02:36 -070019<div id="qv-wrapper">
20 <div id="qv">
21 <h2>In this document</h2>
22 <ol id="auto-toc">
23 </ol>
24 </div>
25</div>
26
27<p>
28This article explains how the Android's audio system attempts to avoid
Clay Murphyc28f2372013-09-25 16:13:40 -070029priority inversion, as of the Android 4.1 release,
Glenn Kasten98afa532013-04-15 14:02:36 -070030and highlights techniques that you can use too.
31</p>
32
33<p>
34These techniques may be useful to developers of high-performance
35audio apps, OEMs, and SoC providers who are implementing an audio
Clay Murphyc28f2372013-09-25 16:13:40 -070036HAL. Please note implementing these techniques is not
Glenn Kasten98afa532013-04-15 14:02:36 -070037guaranteed to prevent glitches or other failures, particularly if
38used outside of the audio context.
Clay Murphyc28f2372013-09-25 16:13:40 -070039Your results may vary, and you should conduct your own
Glenn Kasten98afa532013-04-15 14:02:36 -070040evaluation and testing.
41</p>
42
43<h2 id="background">Background</h2>
44
45<p>
Clay Murphyc28f2372013-09-25 16:13:40 -070046The Android AudioFlinger audio server and AudioTrack/AudioRecord
Glenn Kasten98afa532013-04-15 14:02:36 -070047client implementation are being re-architected to reduce latency.
Clay Murphyc28f2372013-09-25 16:13:40 -070048This work started in Android 4.1, continued in 4.2 and 4.3, and now more
49improvements exist in version 4.4.
Glenn Kasten98afa532013-04-15 14:02:36 -070050</p>
51
52<p>
Clay Murphyc28f2372013-09-25 16:13:40 -070053To achieve this lower latency, many changes were needed throughout the system. One
54important change is to assign CPU resources to time-critical
Glenn Kasten98afa532013-04-15 14:02:36 -070055threads with a more predictable scheduling policy. Reliable scheduling
Clay Murphyc28f2372013-09-25 16:13:40 -070056allows the audio buffer sizes and counts to be reduced while still
Glenn Kasten98afa532013-04-15 14:02:36 -070057avoiding artifacts due to underruns.
58</p>
59
Clay Murphy5d83ab42014-09-09 17:29:09 -070060<h2 id="priorityInversion">Priority inversion</h2>
Glenn Kasten98afa532013-04-15 14:02:36 -070061
62<p>
63<a href="http://en.wikipedia.org/wiki/Priority_inversion">Priority inversion</a>
64is a classic failure mode of real-time systems,
65where a higher-priority task is blocked for an unbounded time waiting
Clay Murphy5d83ab42014-09-09 17:29:09 -070066for a lower-priority task to release a resource such as (shared
67state protected by) a
Glenn Kasten98afa532013-04-15 14:02:36 -070068<a href="http://en.wikipedia.org/wiki/Mutual_exclusion">mutex</a>.
69</p>
70
71<p>
72In an audio system, priority inversion typically manifests as a
73<a href="http://en.wikipedia.org/wiki/Glitch">glitch</a>
74(click, pop, dropout),
75<a href="http://en.wikipedia.org/wiki/Max_Headroom_(character)">repeated audio</a>
76when circular buffers
77are used, or delay in responding to a command.
78</p>
79
80<p>
81In the Android audio implementation, priority inversion is most
Clay Murphy5d83ab42014-09-09 17:29:09 -070082likely to occur in these places. And so you should focus your attention here:
Glenn Kasten98afa532013-04-15 14:02:36 -070083</p>
84
85<ul>
86
87<li>
88between normal mixer thread and fast mixer thread in AudioFlinger
89</li>
90
91<li>
92between application callback thread for a fast AudioTrack and
93fast mixer thread (they both have elevated priority, but slightly
94different priorities)
95</li>
96
97<li>
Clay Murphyc28f2372013-09-25 16:13:40 -070098within the audio Hardware Abstraction Layer (HAL) implementation, e.g. for telephony or echo cancellation
Glenn Kasten98afa532013-04-15 14:02:36 -070099</li>
100
101<li>
102within the audio driver in kernel
103</li>
104
105<li>
106between AudioTrack callback thread and other app threads (this is out of our control)
107</li>
108
109</ul>
110
111<p>
112As of this writing, reduced latency for AudioRecord is planned but
113not yet implemented. The likely priority inversion spots will be
114similar to those for AudioTrack.
115</p>
116
Clay Murphy5d83ab42014-09-09 17:29:09 -0700117<h2 id="commonSolutions">Common solutions</h2>
Glenn Kasten98afa532013-04-15 14:02:36 -0700118
119<p>
Clay Murphy5d83ab42014-09-09 17:29:09 -0700120The typical solutions include:
Glenn Kasten98afa532013-04-15 14:02:36 -0700121</p>
122
123<ul>
124
125<li>
126disabling interrupts
127</li>
128
129<li>
130priority inheritance mutexes
131</li>
132
133</ul>
134
135<p>
136Disabling interrupts is not feasible in Linux user space, and does
Clay Murphyc28f2372013-09-25 16:13:40 -0700137not work for Symmetric Multi-Processors (SMP).
Glenn Kasten98afa532013-04-15 14:02:36 -0700138</p>
139
140
141<p>
142Priority inheritance
143<a href="http://en.wikipedia.org/wiki/Futex">futexes</a>
144(fast user-space mutexes) are available
145in Linux kernel, but are not currently exposed by the Android C
146runtime library
147<a href="http://en.wikipedia.org/wiki/Bionic_(software)">Bionic</a>.
Clay Murphy5d83ab42014-09-09 17:29:09 -0700148They are not used in the audio system because they are relatively heavyweight,
149and because they rely on a trusted client.
Glenn Kasten98afa532013-04-15 14:02:36 -0700150</p>
151
152<h2 id="androidTechniques">Techniques used by Android</h2>
153
154<p>
Clay Murphy5d83ab42014-09-09 17:29:09 -0700155Experiments started with "try lock" and lock with timeout. These are
Glenn Kasten98afa532013-04-15 14:02:36 -0700156non-blocking and bounded blocking variants of the mutex lock
Clay Murphy5d83ab42014-09-09 17:29:09 -0700157operation. Try lock and lock with timeout worked fairly well but were
158susceptible to a couple of obscure failure modes: the
Glenn Kasten98afa532013-04-15 14:02:36 -0700159server was not guaranteed to be able to access the shared state if
160the client happened to be busy, and the cumulative timeout could
161be too long if there was a long sequence of unrelated locks that
162all timed out.
163</p>
164
165
166<p>
167We also use
168<a href="http://en.wikipedia.org/wiki/Linearizability">atomic operations</a>
169such as:
170</p>
171
172<ul>
173<li>increment</li>
174<li>bitwise "or"</li>
175<li>bitwise "and"</li>
176</ul>
177
178<p>
Clay Murphyc28f2372013-09-25 16:13:40 -0700179All of these return the previous value and include the necessary
Glenn Kasten98afa532013-04-15 14:02:36 -0700180SMP barriers. The disadvantage is they can require unbounded retries.
181In practice, we've found that the retries are not a problem.
182</p>
183
Clay Murphy5d83ab42014-09-09 17:29:09 -0700184<p class="note"><strong>Note:</strong> Atomic operations and their interactions with memory barriers
185are notoriously badly misunderstood and used incorrectly. We include these methods
186here for completeness but recommend you also read the article
Glenn Kasten98afa532013-04-15 14:02:36 -0700187<a href="https://developer.android.com/training/articles/smp.html">
188SMP Primer for Android</a>
189for further information.
190</p>
191
192<p>
193We still have and use most of the above tools, and have recently
194added these techniques:
195</p>
196
197<ul>
198
199<li>
200Use non-blocking single-reader single-writer
201<a href="http://en.wikipedia.org/wiki/Circular_buffer">FIFO queues</a>
202for data.
203</li>
204
205<li>
206Try to
207<i>copy</i>
208state rather than
209<i>share</i>
210state between high- and
211low-priority modules.
212</li>
213
214<li>
215When state does need to be shared, limit the state to the
216maximum-size
217<a href="http://en.wikipedia.org/wiki/Word_(computer_architecture)">word</a>
Clay Murphyc28f2372013-09-25 16:13:40 -0700218that can be accessed atomically in one-bus operation
Glenn Kasten98afa532013-04-15 14:02:36 -0700219without retries.
220</li>
221
222<li>
223For complex multi-word state, use a state queue. A state queue
224is basically just a non-blocking single-reader single-writer FIFO
225queue used for state rather than data, except the writer collapses
226adjacent pushes into a single push.
227</li>
228
229<li>
230Pay attention to
231<a href="http://en.wikipedia.org/wiki/Memory_barrier">memory barriers</a>
232for SMP correctness.
233</li>
234
235<li>
236<a href="http://en.wikipedia.org/wiki/Trust,_but_verify">Trust, but verify</a>.
237When sharing
238<i>state</i>
239between processes, don't
240assume that the state is well-formed. For example, check that indices
241are within bounds. This verification isn't needed between threads
242in the same process, between mutual trusting processes (which
243typically have the same UID). It's also unnecessary for shared
244<i>data</i>
245such as PCM audio where a corruption is inconsequential.
246</li>
247
248</ul>
249
Clay Murphy5d83ab42014-09-09 17:29:09 -0700250<h2 id="nonBlockingAlgorithms">Non-blocking algorithms</h2>
Glenn Kasten98afa532013-04-15 14:02:36 -0700251
252<p>
253<a href="http://en.wikipedia.org/wiki/Non-blocking_algorithm">Non-blocking algorithms</a>
254have been a subject of much recent study.
255But with the exception of single-reader single-writer FIFO queues,
256we've found them to be complex and error-prone.
257</p>
258
259<p>
Clay Murphyc28f2372013-09-25 16:13:40 -0700260Starting in Android 4.2, you can find our non-blocking,
Glenn Kasten98afa532013-04-15 14:02:36 -0700261single-reader/writer classes in these locations:
262</p>
263
264<ul>
265
266<li>
267frameworks/av/include/media/nbaio/
268</li>
269
270<li>
271frameworks/av/media/libnbaio/
272</li>
273
274<li>
275frameworks/av/services/audioflinger/StateQueue*
276</li>
277
278</ul>
279
280<p>
281These were designed specifically for AudioFlinger and are not
282general-purpose. Non-blocking algorithms are notorious for being
Clay Murphyc28f2372013-09-25 16:13:40 -0700283difficult to debug. You can look at this code as a model. But be
Glenn Kasten98afa532013-04-15 14:02:36 -0700284aware there may be bugs, and the classes are not guaranteed to be
285suitable for other purposes.
286</p>
287
288<p>
Clay Murphy5d83ab42014-09-09 17:29:09 -0700289For developers, some of the sample OpenSL ES application code should be updated to
290use non-blocking algorithms or reference a non-Android open source library.
Glenn Kasten98afa532013-04-15 14:02:36 -0700291</p>
292
293<h2 id="tools">Tools</h2>
294
295<p>
296To the best of our knowledge, there are no automatic tools for
297finding priority inversion, especially before it happens. Some
298research static code analysis tools are capable of finding priority
299inversions if able to access the entire codebase. Of course, if
300arbitrary user code is involved (as it is here for the application)
301or is a large codebase (as for the Linux kernel and device drivers),
302static analysis may be impractical. The most important thing is to
303read the code very carefully and get a good grasp on the entire
304system and the interactions. Tools such as
305<a href="http://developer.android.com/tools/help/systrace.html">systrace</a>
306and
307<code>ps -t -p</code>
308are useful for seeing priority inversion after it occurs, but do
309not tell you in advance.
310</p>
311
Clay Murphy5d83ab42014-09-09 17:29:09 -0700312<h2 id="aFinalWord">A final word</h2>
Glenn Kasten98afa532013-04-15 14:02:36 -0700313
314<p>
315After all of this discussion, don't be afraid of mutexes. Mutexes
316are your friend for ordinary use, when used and implemented correctly
317in ordinary non-time-critical use cases. But between high- and
318low-priority tasks and in time-sensitive systems mutexes are more
319likely to cause trouble.
320</p>
321