blob: ea8d7b4e53f05c01116f57f45d28da3e9943bfd8
1 | Static Keys |
2 | ----------- |
3 | |
4 | DEPRECATED API: |
5 | |
6 | The use of 'struct static_key' directly, is now DEPRECATED. In addition |
7 | static_key_{true,false}() is also DEPRECATED. IE DO NOT use the following: |
8 | |
9 | struct static_key false = STATIC_KEY_INIT_FALSE; |
10 | struct static_key true = STATIC_KEY_INIT_TRUE; |
11 | static_key_true() |
12 | static_key_false() |
13 | |
14 | The updated API replacements are: |
15 | |
16 | DEFINE_STATIC_KEY_TRUE(key); |
17 | DEFINE_STATIC_KEY_FALSE(key); |
18 | DEFINE_STATIC_KEY_ARRAY_TRUE(keys, count); |
19 | DEFINE_STATIC_KEY_ARRAY_FALSE(keys, count); |
20 | static_branch_likely() |
21 | static_branch_unlikely() |
22 | |
23 | 0) Abstract |
24 | |
25 | Static keys allows the inclusion of seldom used features in |
26 | performance-sensitive fast-path kernel code, via a GCC feature and a code |
27 | patching technique. A quick example: |
28 | |
29 | DEFINE_STATIC_KEY_FALSE(key); |
30 | |
31 | ... |
32 | |
33 | if (static_branch_unlikely(&key)) |
34 | do unlikely code |
35 | else |
36 | do likely code |
37 | |
38 | ... |
39 | static_branch_enable(&key); |
40 | ... |
41 | static_branch_disable(&key); |
42 | ... |
43 | |
44 | The static_branch_unlikely() branch will be generated into the code with as little |
45 | impact to the likely code path as possible. |
46 | |
47 | |
48 | 1) Motivation |
49 | |
50 | |
51 | Currently, tracepoints are implemented using a conditional branch. The |
52 | conditional check requires checking a global variable for each tracepoint. |
53 | Although the overhead of this check is small, it increases when the memory |
54 | cache comes under pressure (memory cache lines for these global variables may |
55 | be shared with other memory accesses). As we increase the number of tracepoints |
56 | in the kernel this overhead may become more of an issue. In addition, |
57 | tracepoints are often dormant (disabled) and provide no direct kernel |
58 | functionality. Thus, it is highly desirable to reduce their impact as much as |
59 | possible. Although tracepoints are the original motivation for this work, other |
60 | kernel code paths should be able to make use of the static keys facility. |
61 | |
62 | |
63 | 2) Solution |
64 | |
65 | |
66 | gcc (v4.5) adds a new 'asm goto' statement that allows branching to a label: |
67 | |
68 | http://gcc.gnu.org/ml/gcc-patches/2009-07/msg01556.html |
69 | |
70 | Using the 'asm goto', we can create branches that are either taken or not taken |
71 | by default, without the need to check memory. Then, at run-time, we can patch |
72 | the branch site to change the branch direction. |
73 | |
74 | For example, if we have a simple branch that is disabled by default: |
75 | |
76 | if (static_branch_unlikely(&key)) |
77 | printk("I am the true branch\n"); |
78 | |
79 | Thus, by default the 'printk' will not be emitted. And the code generated will |
80 | consist of a single atomic 'no-op' instruction (5 bytes on x86), in the |
81 | straight-line code path. When the branch is 'flipped', we will patch the |
82 | 'no-op' in the straight-line codepath with a 'jump' instruction to the |
83 | out-of-line true branch. Thus, changing branch direction is expensive but |
84 | branch selection is basically 'free'. That is the basic tradeoff of this |
85 | optimization. |
86 | |
87 | This lowlevel patching mechanism is called 'jump label patching', and it gives |
88 | the basis for the static keys facility. |
89 | |
90 | 3) Static key label API, usage and examples: |
91 | |
92 | |
93 | In order to make use of this optimization you must first define a key: |
94 | |
95 | DEFINE_STATIC_KEY_TRUE(key); |
96 | |
97 | or: |
98 | |
99 | DEFINE_STATIC_KEY_FALSE(key); |
100 | |
101 | |
102 | The key must be global, that is, it can't be allocated on the stack or dynamically |
103 | allocated at run-time. |
104 | |
105 | The key is then used in code as: |
106 | |
107 | if (static_branch_unlikely(&key)) |
108 | do unlikely code |
109 | else |
110 | do likely code |
111 | |
112 | Or: |
113 | |
114 | if (static_branch_likely(&key)) |
115 | do likely code |
116 | else |
117 | do unlikely code |
118 | |
119 | Keys defined via DEFINE_STATIC_KEY_TRUE(), or DEFINE_STATIC_KEY_FALSE, may |
120 | be used in either static_branch_likely() or static_branch_unlikely() |
121 | statemnts. |
122 | |
123 | Branch(es) can be set true via: |
124 | |
125 | static_branch_enable(&key); |
126 | |
127 | or false via: |
128 | |
129 | static_branch_disable(&key); |
130 | |
131 | The branch(es) can then be switched via reference counts: |
132 | |
133 | static_branch_inc(&key); |
134 | ... |
135 | static_branch_dec(&key); |
136 | |
137 | Thus, 'static_branch_inc()' means 'make the branch true', and |
138 | 'static_branch_dec()' means 'make the branch false' with appropriate |
139 | reference counting. For example, if the key is initialized true, a |
140 | static_branch_dec(), will switch the branch to false. And a subsequent |
141 | static_branch_inc(), will change the branch back to true. Likewise, if the |
142 | key is initialized false, a 'static_branch_inc()', will change the branch to |
143 | true. And then a 'static_branch_dec()', will again make the branch false. |
144 | |
145 | Where an array of keys is required, it can be defined as: |
146 | |
147 | DEFINE_STATIC_KEY_ARRAY_TRUE(keys, count); |
148 | |
149 | or: |
150 | |
151 | DEFINE_STATIC_KEY_ARRAY_FALSE(keys, count); |
152 | |
153 | 4) Architecture level code patching interface, 'jump labels' |
154 | |
155 | |
156 | There are a few functions and macros that architectures must implement in order |
157 | to take advantage of this optimization. If there is no architecture support, we |
158 | simply fall back to a traditional, load, test, and jump sequence. |
159 | |
160 | * select HAVE_ARCH_JUMP_LABEL, see: arch/x86/Kconfig |
161 | |
162 | * #define JUMP_LABEL_NOP_SIZE, see: arch/x86/include/asm/jump_label.h |
163 | |
164 | * __always_inline bool arch_static_branch(struct static_key *key, bool branch), see: |
165 | arch/x86/include/asm/jump_label.h |
166 | |
167 | * __always_inline bool arch_static_branch_jump(struct static_key *key, bool branch), |
168 | see: arch/x86/include/asm/jump_label.h |
169 | |
170 | * void arch_jump_label_transform(struct jump_entry *entry, enum jump_label_type type), |
171 | see: arch/x86/kernel/jump_label.c |
172 | |
173 | * __init_or_module void arch_jump_label_transform_static(struct jump_entry *entry, enum jump_label_type type), |
174 | see: arch/x86/kernel/jump_label.c |
175 | |
176 | |
177 | * struct jump_entry, see: arch/x86/include/asm/jump_label.h |
178 | |
179 | |
180 | 5) Static keys / jump label analysis, results (x86_64): |
181 | |
182 | |
183 | As an example, let's add the following branch to 'getppid()', such that the |
184 | system call now looks like: |
185 | |
186 | SYSCALL_DEFINE0(getppid) |
187 | { |
188 | int pid; |
189 | |
190 | + if (static_branch_unlikely(&key)) |
191 | + printk("I am the true branch\n"); |
192 | |
193 | rcu_read_lock(); |
194 | pid = task_tgid_vnr(rcu_dereference(current->real_parent)); |
195 | rcu_read_unlock(); |
196 | |
197 | return pid; |
198 | } |
199 | |
200 | The resulting instructions with jump labels generated by GCC is: |
201 | |
202 | ffffffff81044290 <sys_getppid>: |
203 | ffffffff81044290: 55 push %rbp |
204 | ffffffff81044291: 48 89 e5 mov %rsp,%rbp |
205 | ffffffff81044294: e9 00 00 00 00 jmpq ffffffff81044299 <sys_getppid+0x9> |
206 | ffffffff81044299: 65 48 8b 04 25 c0 b6 mov %gs:0xb6c0,%rax |
207 | ffffffff810442a0: 00 00 |
208 | ffffffff810442a2: 48 8b 80 80 02 00 00 mov 0x280(%rax),%rax |
209 | ffffffff810442a9: 48 8b 80 b0 02 00 00 mov 0x2b0(%rax),%rax |
210 | ffffffff810442b0: 48 8b b8 e8 02 00 00 mov 0x2e8(%rax),%rdi |
211 | ffffffff810442b7: e8 f4 d9 00 00 callq ffffffff81051cb0 <pid_vnr> |
212 | ffffffff810442bc: 5d pop %rbp |
213 | ffffffff810442bd: 48 98 cltq |
214 | ffffffff810442bf: c3 retq |
215 | ffffffff810442c0: 48 c7 c7 e3 54 98 81 mov $0xffffffff819854e3,%rdi |
216 | ffffffff810442c7: 31 c0 xor %eax,%eax |
217 | ffffffff810442c9: e8 71 13 6d 00 callq ffffffff8171563f <printk> |
218 | ffffffff810442ce: eb c9 jmp ffffffff81044299 <sys_getppid+0x9> |
219 | |
220 | Without the jump label optimization it looks like: |
221 | |
222 | ffffffff810441f0 <sys_getppid>: |
223 | ffffffff810441f0: 8b 05 8a 52 d8 00 mov 0xd8528a(%rip),%eax # ffffffff81dc9480 <key> |
224 | ffffffff810441f6: 55 push %rbp |
225 | ffffffff810441f7: 48 89 e5 mov %rsp,%rbp |
226 | ffffffff810441fa: 85 c0 test %eax,%eax |
227 | ffffffff810441fc: 75 27 jne ffffffff81044225 <sys_getppid+0x35> |
228 | ffffffff810441fe: 65 48 8b 04 25 c0 b6 mov %gs:0xb6c0,%rax |
229 | ffffffff81044205: 00 00 |
230 | ffffffff81044207: 48 8b 80 80 02 00 00 mov 0x280(%rax),%rax |
231 | ffffffff8104420e: 48 8b 80 b0 02 00 00 mov 0x2b0(%rax),%rax |
232 | ffffffff81044215: 48 8b b8 e8 02 00 00 mov 0x2e8(%rax),%rdi |
233 | ffffffff8104421c: e8 2f da 00 00 callq ffffffff81051c50 <pid_vnr> |
234 | ffffffff81044221: 5d pop %rbp |
235 | ffffffff81044222: 48 98 cltq |
236 | ffffffff81044224: c3 retq |
237 | ffffffff81044225: 48 c7 c7 13 53 98 81 mov $0xffffffff81985313,%rdi |
238 | ffffffff8104422c: 31 c0 xor %eax,%eax |
239 | ffffffff8104422e: e8 60 0f 6d 00 callq ffffffff81715193 <printk> |
240 | ffffffff81044233: eb c9 jmp ffffffff810441fe <sys_getppid+0xe> |
241 | ffffffff81044235: 66 66 2e 0f 1f 84 00 data32 nopw %cs:0x0(%rax,%rax,1) |
242 | ffffffff8104423c: 00 00 00 00 |
243 | |
244 | Thus, the disable jump label case adds a 'mov', 'test' and 'jne' instruction |
245 | vs. the jump label case just has a 'no-op' or 'jmp 0'. (The jmp 0, is patched |
246 | to a 5 byte atomic no-op instruction at boot-time.) Thus, the disabled jump |
247 | label case adds: |
248 | |
249 | 6 (mov) + 2 (test) + 2 (jne) = 10 - 5 (5 byte jump 0) = 5 addition bytes. |
250 | |
251 | If we then include the padding bytes, the jump label code saves, 16 total bytes |
252 | of instruction memory for this small function. In this case the non-jump label |
253 | function is 80 bytes long. Thus, we have saved 20% of the instruction |
254 | footprint. We can in fact improve this even further, since the 5-byte no-op |
255 | really can be a 2-byte no-op since we can reach the branch with a 2-byte jmp. |
256 | However, we have not yet implemented optimal no-op sizes (they are currently |
257 | hard-coded). |
258 | |
259 | Since there are a number of static key API uses in the scheduler paths, |
260 | 'pipe-test' (also known as 'perf bench sched pipe') can be used to show the |
261 | performance improvement. Testing done on 3.3.0-rc2: |
262 | |
263 | jump label disabled: |
264 | |
265 | Performance counter stats for 'bash -c /tmp/pipe-test' (50 runs): |
266 | |
267 | 855.700314 task-clock # 0.534 CPUs utilized ( +- 0.11% ) |
268 | 200,003 context-switches # 0.234 M/sec ( +- 0.00% ) |
269 | 0 CPU-migrations # 0.000 M/sec ( +- 39.58% ) |
270 | 487 page-faults # 0.001 M/sec ( +- 0.02% ) |
271 | 1,474,374,262 cycles # 1.723 GHz ( +- 0.17% ) |
272 | <not supported> stalled-cycles-frontend |
273 | <not supported> stalled-cycles-backend |
274 | 1,178,049,567 instructions # 0.80 insns per cycle ( +- 0.06% ) |
275 | 208,368,926 branches # 243.507 M/sec ( +- 0.06% ) |
276 | 5,569,188 branch-misses # 2.67% of all branches ( +- 0.54% ) |
277 | |
278 | 1.601607384 seconds time elapsed ( +- 0.07% ) |
279 | |
280 | jump label enabled: |
281 | |
282 | Performance counter stats for 'bash -c /tmp/pipe-test' (50 runs): |
283 | |
284 | 841.043185 task-clock # 0.533 CPUs utilized ( +- 0.12% ) |
285 | 200,004 context-switches # 0.238 M/sec ( +- 0.00% ) |
286 | 0 CPU-migrations # 0.000 M/sec ( +- 40.87% ) |
287 | 487 page-faults # 0.001 M/sec ( +- 0.05% ) |
288 | 1,432,559,428 cycles # 1.703 GHz ( +- 0.18% ) |
289 | <not supported> stalled-cycles-frontend |
290 | <not supported> stalled-cycles-backend |
291 | 1,175,363,994 instructions # 0.82 insns per cycle ( +- 0.04% ) |
292 | 206,859,359 branches # 245.956 M/sec ( +- 0.04% ) |
293 | 4,884,119 branch-misses # 2.36% of all branches ( +- 0.85% ) |
294 | |
295 | 1.579384366 seconds time elapsed |
296 | |
297 | The percentage of saved branches is .7%, and we've saved 12% on |
298 | 'branch-misses'. This is where we would expect to get the most savings, since |
299 | this optimization is about reducing the number of branches. In addition, we've |
300 | saved .2% on instructions, and 2.8% on cycles and 1.4% on elapsed time. |
301 |