]>
Commit | Line | Data |
---|---|---|
1 | Static Keys | |
2 | ----------- | |
3 | ||
4 | By: Jason Baron <jbaron@redhat.com> | |
5 | ||
6 | 0) Abstract | |
7 | ||
8 | Static keys allows the inclusion of seldom used features in | |
9 | performance-sensitive fast-path kernel code, via a GCC feature and a code | |
10 | patching technique. A quick example: | |
11 | ||
12 | struct static_key key = STATIC_KEY_INIT_FALSE; | |
13 | ||
14 | ... | |
15 | ||
16 | if (static_key_false(&key)) | |
17 | do unlikely code | |
18 | else | |
19 | do likely code | |
20 | ||
21 | ... | |
22 | static_key_slow_inc(); | |
23 | ... | |
24 | static_key_slow_inc(); | |
25 | ... | |
26 | ||
27 | The static_key_false() branch will be generated into the code with as little | |
28 | impact to the likely code path as possible. | |
29 | ||
30 | ||
31 | 1) Motivation | |
32 | ||
33 | ||
34 | Currently, tracepoints are implemented using a conditional branch. The | |
35 | conditional check requires checking a global variable for each tracepoint. | |
36 | Although the overhead of this check is small, it increases when the memory | |
37 | cache comes under pressure (memory cache lines for these global variables may | |
38 | be shared with other memory accesses). As we increase the number of tracepoints | |
39 | in the kernel this overhead may become more of an issue. In addition, | |
40 | tracepoints are often dormant (disabled) and provide no direct kernel | |
41 | functionality. Thus, it is highly desirable to reduce their impact as much as | |
42 | possible. Although tracepoints are the original motivation for this work, other | |
43 | kernel code paths should be able to make use of the static keys facility. | |
44 | ||
45 | ||
46 | 2) Solution | |
47 | ||
48 | ||
49 | gcc (v4.5) adds a new 'asm goto' statement that allows branching to a label: | |
50 | ||
51 | http://gcc.gnu.org/ml/gcc-patches/2009-07/msg01556.html | |
52 | ||
53 | Using the 'asm goto', we can create branches that are either taken or not taken | |
54 | by default, without the need to check memory. Then, at run-time, we can patch | |
55 | the branch site to change the branch direction. | |
56 | ||
57 | For example, if we have a simple branch that is disabled by default: | |
58 | ||
59 | if (static_key_false(&key)) | |
60 | printk("I am the true branch\n"); | |
61 | ||
62 | Thus, by default the 'printk' will not be emitted. And the code generated will | |
63 | consist of a single atomic 'no-op' instruction (5 bytes on x86), in the | |
64 | straight-line code path. When the branch is 'flipped', we will patch the | |
65 | 'no-op' in the straight-line codepath with a 'jump' instruction to the | |
66 | out-of-line true branch. Thus, changing branch direction is expensive but | |
67 | branch selection is basically 'free'. That is the basic tradeoff of this | |
68 | optimization. | |
69 | ||
70 | This lowlevel patching mechanism is called 'jump label patching', and it gives | |
71 | the basis for the static keys facility. | |
72 | ||
73 | 3) Static key label API, usage and examples: | |
74 | ||
75 | ||
76 | In order to make use of this optimization you must first define a key: | |
77 | ||
78 | struct static_key key; | |
79 | ||
80 | Which is initialized as: | |
81 | ||
82 | struct static_key key = STATIC_KEY_INIT_TRUE; | |
83 | ||
84 | or: | |
85 | ||
86 | struct static_key key = STATIC_KEY_INIT_FALSE; | |
87 | ||
88 | If the key is not initialized, it is default false. The 'struct static_key', | |
89 | must be a 'global'. That is, it can't be allocated on the stack or dynamically | |
90 | allocated at run-time. | |
91 | ||
92 | The key is then used in code as: | |
93 | ||
94 | if (static_key_false(&key)) | |
95 | do unlikely code | |
96 | else | |
97 | do likely code | |
98 | ||
99 | Or: | |
100 | ||
101 | if (static_key_true(&key)) | |
102 | do likely code | |
103 | else | |
104 | do unlikely code | |
105 | ||
106 | A key that is initialized via 'STATIC_KEY_INIT_FALSE', must be used in a | |
107 | 'static_key_false()' construct. Likewise, a key initialized via | |
108 | 'STATIC_KEY_INIT_TRUE' must be used in a 'static_key_true()' construct. A | |
109 | single key can be used in many branches, but all the branches must match the | |
110 | way that the key has been initialized. | |
111 | ||
112 | The branch(es) can then be switched via: | |
113 | ||
114 | static_key_slow_inc(&key); | |
115 | ... | |
116 | static_key_slow_dec(&key); | |
117 | ||
118 | Thus, 'static_key_slow_inc()' means 'make the branch true', and | |
119 | 'static_key_slow_dec()' means 'make the branch false' with appropriate | |
120 | reference counting. For example, if the key is initialized true, a | |
121 | static_key_slow_dec(), will switch the branch to false. And a subsequent | |
122 | static_key_slow_inc(), will change the branch back to true. Likewise, if the | |
123 | key is initialized false, a 'static_key_slow_inc()', will change the branch to | |
124 | true. And then a 'static_key_slow_dec()', will again make the branch false. | |
125 | ||
126 | An example usage in the kernel is the implementation of tracepoints: | |
127 | ||
128 | static inline void trace_##name(proto) \ | |
129 | { \ | |
130 | if (static_key_false(&__tracepoint_##name.key)) \ | |
131 | __DO_TRACE(&__tracepoint_##name, \ | |
132 | TP_PROTO(data_proto), \ | |
133 | TP_ARGS(data_args), \ | |
134 | TP_CONDITION(cond)); \ | |
135 | } | |
136 | ||
137 | Tracepoints are disabled by default, and can be placed in performance critical | |
138 | pieces of the kernel. Thus, by using a static key, the tracepoints can have | |
139 | absolutely minimal impact when not in use. | |
140 | ||
141 | ||
142 | 4) Architecture level code patching interface, 'jump labels' | |
143 | ||
144 | ||
145 | There are a few functions and macros that architectures must implement in order | |
146 | to take advantage of this optimization. If there is no architecture support, we | |
147 | simply fall back to a traditional, load, test, and jump sequence. | |
148 | ||
149 | * select HAVE_ARCH_JUMP_LABEL, see: arch/x86/Kconfig | |
150 | ||
151 | * #define JUMP_LABEL_NOP_SIZE, see: arch/x86/include/asm/jump_label.h | |
152 | ||
153 | * __always_inline bool arch_static_branch(struct static_key *key), see: | |
154 | arch/x86/include/asm/jump_label.h | |
155 | ||
156 | * void arch_jump_label_transform(struct jump_entry *entry, enum jump_label_type type), | |
157 | see: arch/x86/kernel/jump_label.c | |
158 | ||
159 | * __init_or_module void arch_jump_label_transform_static(struct jump_entry *entry, enum jump_label_type type), | |
160 | see: arch/x86/kernel/jump_label.c | |
161 | ||
162 | ||
163 | * struct jump_entry, see: arch/x86/include/asm/jump_label.h | |
164 | ||
165 | ||
166 | 5) Static keys / jump label analysis, results (x86_64): | |
167 | ||
168 | ||
169 | As an example, let's add the following branch to 'getppid()', such that the | |
170 | system call now looks like: | |
171 | ||
172 | SYSCALL_DEFINE0(getppid) | |
173 | { | |
174 | int pid; | |
175 | ||
176 | + if (static_key_false(&key)) | |
177 | + printk("I am the true branch\n"); | |
178 | ||
179 | rcu_read_lock(); | |
180 | pid = task_tgid_vnr(rcu_dereference(current->real_parent)); | |
181 | rcu_read_unlock(); | |
182 | ||
183 | return pid; | |
184 | } | |
185 | ||
186 | The resulting instructions with jump labels generated by GCC is: | |
187 | ||
188 | ffffffff81044290 <sys_getppid>: | |
189 | ffffffff81044290: 55 push %rbp | |
190 | ffffffff81044291: 48 89 e5 mov %rsp,%rbp | |
191 | ffffffff81044294: e9 00 00 00 00 jmpq ffffffff81044299 <sys_getppid+0x9> | |
192 | ffffffff81044299: 65 48 8b 04 25 c0 b6 mov %gs:0xb6c0,%rax | |
193 | ffffffff810442a0: 00 00 | |
194 | ffffffff810442a2: 48 8b 80 80 02 00 00 mov 0x280(%rax),%rax | |
195 | ffffffff810442a9: 48 8b 80 b0 02 00 00 mov 0x2b0(%rax),%rax | |
196 | ffffffff810442b0: 48 8b b8 e8 02 00 00 mov 0x2e8(%rax),%rdi | |
197 | ffffffff810442b7: e8 f4 d9 00 00 callq ffffffff81051cb0 <pid_vnr> | |
198 | ffffffff810442bc: 5d pop %rbp | |
199 | ffffffff810442bd: 48 98 cltq | |
200 | ffffffff810442bf: c3 retq | |
201 | ffffffff810442c0: 48 c7 c7 e3 54 98 81 mov $0xffffffff819854e3,%rdi | |
202 | ffffffff810442c7: 31 c0 xor %eax,%eax | |
203 | ffffffff810442c9: e8 71 13 6d 00 callq ffffffff8171563f <printk> | |
204 | ffffffff810442ce: eb c9 jmp ffffffff81044299 <sys_getppid+0x9> | |
205 | ||
206 | Without the jump label optimization it looks like: | |
207 | ||
208 | ffffffff810441f0 <sys_getppid>: | |
209 | ffffffff810441f0: 8b 05 8a 52 d8 00 mov 0xd8528a(%rip),%eax # ffffffff81dc9480 <key> | |
210 | ffffffff810441f6: 55 push %rbp | |
211 | ffffffff810441f7: 48 89 e5 mov %rsp,%rbp | |
212 | ffffffff810441fa: 85 c0 test %eax,%eax | |
213 | ffffffff810441fc: 75 27 jne ffffffff81044225 <sys_getppid+0x35> | |
214 | ffffffff810441fe: 65 48 8b 04 25 c0 b6 mov %gs:0xb6c0,%rax | |
215 | ffffffff81044205: 00 00 | |
216 | ffffffff81044207: 48 8b 80 80 02 00 00 mov 0x280(%rax),%rax | |
217 | ffffffff8104420e: 48 8b 80 b0 02 00 00 mov 0x2b0(%rax),%rax | |
218 | ffffffff81044215: 48 8b b8 e8 02 00 00 mov 0x2e8(%rax),%rdi | |
219 | ffffffff8104421c: e8 2f da 00 00 callq ffffffff81051c50 <pid_vnr> | |
220 | ffffffff81044221: 5d pop %rbp | |
221 | ffffffff81044222: 48 98 cltq | |
222 | ffffffff81044224: c3 retq | |
223 | ffffffff81044225: 48 c7 c7 13 53 98 81 mov $0xffffffff81985313,%rdi | |
224 | ffffffff8104422c: 31 c0 xor %eax,%eax | |
225 | ffffffff8104422e: e8 60 0f 6d 00 callq ffffffff81715193 <printk> | |
226 | ffffffff81044233: eb c9 jmp ffffffff810441fe <sys_getppid+0xe> | |
227 | ffffffff81044235: 66 66 2e 0f 1f 84 00 data32 nopw %cs:0x0(%rax,%rax,1) | |
228 | ffffffff8104423c: 00 00 00 00 | |
229 | ||
230 | Thus, the disable jump label case adds a 'mov', 'test' and 'jne' instruction | |
231 | vs. the jump label case just has a 'no-op' or 'jmp 0'. (The jmp 0, is patched | |
232 | to a 5 byte atomic no-op instruction at boot-time.) Thus, the disabled jump | |
233 | label case adds: | |
234 | ||
235 | 6 (mov) + 2 (test) + 2 (jne) = 10 - 5 (5 byte jump 0) = 5 addition bytes. | |
236 | ||
237 | If we then include the padding bytes, the jump label code saves, 16 total bytes | |
238 | of instruction memory for this small function. In this case the non-jump label | |
239 | function is 80 bytes long. Thus, we have saved 20% of the instruction | |
240 | footprint. We can in fact improve this even further, since the 5-byte no-op | |
241 | really can be a 2-byte no-op since we can reach the branch with a 2-byte jmp. | |
242 | However, we have not yet implemented optimal no-op sizes (they are currently | |
243 | hard-coded). | |
244 | ||
245 | Since there are a number of static key API uses in the scheduler paths, | |
246 | 'pipe-test' (also known as 'perf bench sched pipe') can be used to show the | |
247 | performance improvement. Testing done on 3.3.0-rc2: | |
248 | ||
249 | jump label disabled: | |
250 | ||
251 | Performance counter stats for 'bash -c /tmp/pipe-test' (50 runs): | |
252 | ||
253 | 855.700314 task-clock # 0.534 CPUs utilized ( +- 0.11% ) | |
254 | 200,003 context-switches # 0.234 M/sec ( +- 0.00% ) | |
255 | 0 CPU-migrations # 0.000 M/sec ( +- 39.58% ) | |
256 | 487 page-faults # 0.001 M/sec ( +- 0.02% ) | |
257 | 1,474,374,262 cycles # 1.723 GHz ( +- 0.17% ) | |
258 | <not supported> stalled-cycles-frontend | |
259 | <not supported> stalled-cycles-backend | |
260 | 1,178,049,567 instructions # 0.80 insns per cycle ( +- 0.06% ) | |
261 | 208,368,926 branches # 243.507 M/sec ( +- 0.06% ) | |
262 | 5,569,188 branch-misses # 2.67% of all branches ( +- 0.54% ) | |
263 | ||
264 | 1.601607384 seconds time elapsed ( +- 0.07% ) | |
265 | ||
266 | jump label enabled: | |
267 | ||
268 | Performance counter stats for 'bash -c /tmp/pipe-test' (50 runs): | |
269 | ||
270 | 841.043185 task-clock # 0.533 CPUs utilized ( +- 0.12% ) | |
271 | 200,004 context-switches # 0.238 M/sec ( +- 0.00% ) | |
272 | 0 CPU-migrations # 0.000 M/sec ( +- 40.87% ) | |
273 | 487 page-faults # 0.001 M/sec ( +- 0.05% ) | |
274 | 1,432,559,428 cycles # 1.703 GHz ( +- 0.18% ) | |
275 | <not supported> stalled-cycles-frontend | |
276 | <not supported> stalled-cycles-backend | |
277 | 1,175,363,994 instructions # 0.82 insns per cycle ( +- 0.04% ) | |
278 | 206,859,359 branches # 245.956 M/sec ( +- 0.04% ) | |
279 | 4,884,119 branch-misses # 2.36% of all branches ( +- 0.85% ) | |
280 | ||
281 | 1.579384366 seconds time elapsed | |
282 | ||
283 | The percentage of saved branches is .7%, and we've saved 12% on | |
284 | 'branch-misses'. This is where we would expect to get the most savings, since | |
285 | this optimization is about reducing the number of branches. In addition, we've | |
286 | saved .2% on instructions, and 2.8% on cycles and 1.4% on elapsed time. |