| 1 |
=========== |
| 2 |
Static Keys |
| 3 |
=========== |
| 4 |
|
| 5 |
.. warning:: |
| 6 |
|
| 7 |
DEPRECATED API: |
| 8 |
|
| 9 |
The use of 'struct static_key' directly, is now DEPRECATED. In addition |
| 10 |
static_key_{true,false}() is also DEPRECATED. IE DO NOT use the following:: |
| 11 |
|
| 12 |
struct static_key false = STATIC_KEY_INIT_FALSE; |
| 13 |
struct static_key true = STATIC_KEY_INIT_TRUE; |
| 14 |
static_key_true() |
| 15 |
static_key_false() |
| 16 |
|
| 17 |
The updated API replacements are:: |
| 18 |
|
| 19 |
DEFINE_STATIC_KEY_TRUE(key); |
| 20 |
DEFINE_STATIC_KEY_FALSE(key); |
| 21 |
DEFINE_STATIC_KEY_ARRAY_TRUE(keys, count); |
| 22 |
DEFINE_STATIC_KEY_ARRAY_FALSE(keys, count); |
| 23 |
static_branch_likely() |
| 24 |
static_branch_unlikely() |
| 25 |
|
| 26 |
Abstract |
| 27 |
======== |
| 28 |
|
| 29 |
Static keys allows the inclusion of seldom used features in |
| 30 |
performance-sensitive fast-path kernel code, via a GCC feature and a code |
| 31 |
patching technique. A quick example:: |
| 32 |
|
| 33 |
DEFINE_STATIC_KEY_FALSE(key); |
| 34 |
|
| 35 |
... |
| 36 |
|
| 37 |
if (static_branch_unlikely(&key)) |
| 38 |
do unlikely code |
| 39 |
else |
| 40 |
do likely code |
| 41 |
|
| 42 |
... |
| 43 |
static_branch_enable(&key); |
| 44 |
... |
| 45 |
static_branch_disable(&key); |
| 46 |
... |
| 47 |
|
| 48 |
The static_branch_unlikely() branch will be generated into the code with as little |
| 49 |
impact to the likely code path as possible. |
| 50 |
|
| 51 |
|
| 52 |
Motivation |
| 53 |
========== |
| 54 |
|
| 55 |
|
| 56 |
Currently, tracepoints are implemented using a conditional branch. The |
| 57 |
conditional check requires checking a global variable for each tracepoint. |
| 58 |
Although the overhead of this check is small, it increases when the memory |
| 59 |
cache comes under pressure (memory cache lines for these global variables may |
| 60 |
be shared with other memory accesses). As we increase the number of tracepoints |
| 61 |
in the kernel this overhead may become more of an issue. In addition, |
| 62 |
tracepoints are often dormant (disabled) and provide no direct kernel |
| 63 |
functionality. Thus, it is highly desirable to reduce their impact as much as |
| 64 |
possible. Although tracepoints are the original motivation for this work, other |
| 65 |
kernel code paths should be able to make use of the static keys facility. |
| 66 |
|
| 67 |
|
| 68 |
Solution |
| 69 |
======== |
| 70 |
|
| 71 |
|
| 72 |
gcc (v4.5) adds a new 'asm goto' statement that allows branching to a label: |
| 73 |
|
| 74 |
https://gcc.gnu.org/ml/gcc-patches/2009-07/msg01556.html |
| 75 |
|
| 76 |
Using the 'asm goto', we can create branches that are either taken or not taken |
| 77 |
by default, without the need to check memory. Then, at run-time, we can patch |
| 78 |
the branch site to change the branch direction. |
| 79 |
|
| 80 |
For example, if we have a simple branch that is disabled by default:: |
| 81 |
|
| 82 |
if (static_branch_unlikely(&key)) |
| 83 |
printk("I am the true branch\n"); |
| 84 |
|
| 85 |
Thus, by default the 'printk' will not be emitted. And the code generated will |
| 86 |
consist of a single atomic 'no-op' instruction (5 bytes on x86), in the |
| 87 |
straight-line code path. When the branch is 'flipped', we will patch the |
| 88 |
'no-op' in the straight-line codepath with a 'jump' instruction to the |
| 89 |
out-of-line true branch. Thus, changing branch direction is expensive but |
| 90 |
branch selection is basically 'free'. That is the basic tradeoff of this |
| 91 |
optimization. |
| 92 |
|
| 93 |
This lowlevel patching mechanism is called 'jump label patching', and it gives |
| 94 |
the basis for the static keys facility. |
| 95 |
|
| 96 |
Static key label API, usage and examples |
| 97 |
======================================== |
| 98 |
|
| 99 |
|
| 100 |
In order to make use of this optimization you must first define a key:: |
| 101 |
|
| 102 |
DEFINE_STATIC_KEY_TRUE(key); |
| 103 |
|
| 104 |
or:: |
| 105 |
|
| 106 |
DEFINE_STATIC_KEY_FALSE(key); |
| 107 |
|
| 108 |
|
| 109 |
The key must be global, that is, it can't be allocated on the stack or dynamically |
| 110 |
allocated at run-time. |
| 111 |
|
| 112 |
The key is then used in code as:: |
| 113 |
|
| 114 |
if (static_branch_unlikely(&key)) |
| 115 |
do unlikely code |
| 116 |
else |
| 117 |
do likely code |
| 118 |
|
| 119 |
Or:: |
| 120 |
|
| 121 |
if (static_branch_likely(&key)) |
| 122 |
do likely code |
| 123 |
else |
| 124 |
do unlikely code |
| 125 |
|
| 126 |
Keys defined via DEFINE_STATIC_KEY_TRUE(), or DEFINE_STATIC_KEY_FALSE, may |
| 127 |
be used in either static_branch_likely() or static_branch_unlikely() |
| 128 |
statements. |
| 129 |
|
| 130 |
Branch(es) can be set true via:: |
| 131 |
|
| 132 |
static_branch_enable(&key); |
| 133 |
|
| 134 |
or false via:: |
| 135 |
|
| 136 |
static_branch_disable(&key); |
| 137 |
|
| 138 |
The branch(es) can then be switched via reference counts:: |
| 139 |
|
| 140 |
static_branch_inc(&key); |
| 141 |
... |
| 142 |
static_branch_dec(&key); |
| 143 |
|
| 144 |
Thus, 'static_branch_inc()' means 'make the branch true', and |
| 145 |
'static_branch_dec()' means 'make the branch false' with appropriate |
| 146 |
reference counting. For example, if the key is initialized true, a |
| 147 |
static_branch_dec(), will switch the branch to false. And a subsequent |
| 148 |
static_branch_inc(), will change the branch back to true. Likewise, if the |
| 149 |
key is initialized false, a 'static_branch_inc()', will change the branch to |
| 150 |
true. And then a 'static_branch_dec()', will again make the branch false. |
| 151 |
|
| 152 |
The state and the reference count can be retrieved with 'static_key_enabled()' |
| 153 |
and 'static_key_count()'. In general, if you use these functions, they |
| 154 |
should be protected with the same mutex used around the enable/disable |
| 155 |
or increment/decrement function. |
| 156 |
|
| 157 |
Note that switching branches results in some locks being taken, |
| 158 |
particularly the CPU hotplug lock (in order to avoid races against |
| 159 |
CPUs being brought in the kernel while the kernel is getting |
| 160 |
patched). Calling the static key API from within a hotplug notifier is |
| 161 |
thus a sure deadlock recipe. In order to still allow use of the |
| 162 |
functionality, the following functions are provided: |
| 163 |
|
| 164 |
static_key_enable_cpuslocked() |
| 165 |
static_key_disable_cpuslocked() |
| 166 |
static_branch_enable_cpuslocked() |
| 167 |
static_branch_disable_cpuslocked() |
| 168 |
|
| 169 |
These functions are *not* general purpose, and must only be used when |
| 170 |
you really know that you're in the above context, and no other. |
| 171 |
|
| 172 |
Where an array of keys is required, it can be defined as:: |
| 173 |
|
| 174 |
DEFINE_STATIC_KEY_ARRAY_TRUE(keys, count); |
| 175 |
|
| 176 |
or:: |
| 177 |
|
| 178 |
DEFINE_STATIC_KEY_ARRAY_FALSE(keys, count); |
| 179 |
|
| 180 |
4) Architecture level code patching interface, 'jump labels' |
| 181 |
|
| 182 |
|
| 183 |
There are a few functions and macros that architectures must implement in order |
| 184 |
to take advantage of this optimization. If there is no architecture support, we |
| 185 |
simply fall back to a traditional, load, test, and jump sequence. Also, the |
| 186 |
struct jump_entry table must be at least 4-byte aligned because the |
| 187 |
static_key->entry field makes use of the two least significant bits. |
| 188 |
|
| 189 |
* ``select HAVE_ARCH_JUMP_LABEL``, |
| 190 |
see: arch/x86/Kconfig |
| 191 |
|
| 192 |
* ``#define JUMP_LABEL_NOP_SIZE``, |
| 193 |
see: arch/x86/include/asm/jump_label.h |
| 194 |
|
| 195 |
* ``__always_inline bool arch_static_branch(struct static_key *key, bool branch)``, |
| 196 |
see: arch/x86/include/asm/jump_label.h |
| 197 |
|
| 198 |
* ``__always_inline bool arch_static_branch_jump(struct static_key *key, bool branch)``, |
| 199 |
see: arch/x86/include/asm/jump_label.h |
| 200 |
|
| 201 |
* ``void arch_jump_label_transform(struct jump_entry *entry, enum jump_label_type type)``, |
| 202 |
see: arch/x86/kernel/jump_label.c |
| 203 |
|
| 204 |
* ``struct jump_entry``, |
| 205 |
see: arch/x86/include/asm/jump_label.h |
| 206 |
|
| 207 |
|
| 208 |
5) Static keys / jump label analysis, results (x86_64): |
| 209 |
|
| 210 |
|
| 211 |
As an example, let's add the following branch to 'getppid()', such that the |
| 212 |
system call now looks like:: |
| 213 |
|
| 214 |
SYSCALL_DEFINE0(getppid) |
| 215 |
{ |
| 216 |
int pid; |
| 217 |
|
| 218 |
+ if (static_branch_unlikely(&key)) |
| 219 |
+ printk("I am the true branch\n"); |
| 220 |
|
| 221 |
rcu_read_lock(); |
| 222 |
pid = task_tgid_vnr(rcu_dereference(current->real_parent)); |
| 223 |
rcu_read_unlock(); |
| 224 |
|
| 225 |
return pid; |
| 226 |
} |
| 227 |
|
| 228 |
The resulting instructions with jump labels generated by GCC is:: |
| 229 |
|
| 230 |
ffffffff81044290 <sys_getppid>: |
| 231 |
ffffffff81044290: 55 push %rbp |
| 232 |
ffffffff81044291: 48 89 e5 mov %rsp,%rbp |
| 233 |
ffffffff81044294: e9 00 00 00 00 jmpq ffffffff81044299 <sys_getppid+0x9> |
| 234 |
ffffffff81044299: 65 48 8b 04 25 c0 b6 mov %gs:0xb6c0,%rax |
| 235 |
ffffffff810442a0: 00 00 |
| 236 |
ffffffff810442a2: 48 8b 80 80 02 00 00 mov 0x280(%rax),%rax |
| 237 |
ffffffff810442a9: 48 8b 80 b0 02 00 00 mov 0x2b0(%rax),%rax |
| 238 |
ffffffff810442b0: 48 8b b8 e8 02 00 00 mov 0x2e8(%rax),%rdi |
| 239 |
ffffffff810442b7: e8 f4 d9 00 00 callq ffffffff81051cb0 <pid_vnr> |
| 240 |
ffffffff810442bc: 5d pop %rbp |
| 241 |
ffffffff810442bd: 48 98 cltq |
| 242 |
ffffffff810442bf: c3 retq |
| 243 |
ffffffff810442c0: 48 c7 c7 e3 54 98 81 mov $0xffffffff819854e3,%rdi |
| 244 |
ffffffff810442c7: 31 c0 xor %eax,%eax |
| 245 |
ffffffff810442c9: e8 71 13 6d 00 callq ffffffff8171563f <printk> |
| 246 |
ffffffff810442ce: eb c9 jmp ffffffff81044299 <sys_getppid+0x9> |
| 247 |
|
| 248 |
Without the jump label optimization it looks like:: |
| 249 |
|
| 250 |
ffffffff810441f0 <sys_getppid>: |
| 251 |
ffffffff810441f0: 8b 05 8a 52 d8 00 mov 0xd8528a(%rip),%eax # ffffffff81dc9480 <key> |
| 252 |
ffffffff810441f6: 55 push %rbp |
| 253 |
ffffffff810441f7: 48 89 e5 mov %rsp,%rbp |
| 254 |
ffffffff810441fa: 85 c0 test %eax,%eax |
| 255 |
ffffffff810441fc: 75 27 jne ffffffff81044225 <sys_getppid+0x35> |
| 256 |
ffffffff810441fe: 65 48 8b 04 25 c0 b6 mov %gs:0xb6c0,%rax |
| 257 |
ffffffff81044205: 00 00 |
| 258 |
ffffffff81044207: 48 8b 80 80 02 00 00 mov 0x280(%rax),%rax |
| 259 |
ffffffff8104420e: 48 8b 80 b0 02 00 00 mov 0x2b0(%rax),%rax |
| 260 |
ffffffff81044215: 48 8b b8 e8 02 00 00 mov 0x2e8(%rax),%rdi |
| 261 |
ffffffff8104421c: e8 2f da 00 00 callq ffffffff81051c50 <pid_vnr> |
| 262 |
ffffffff81044221: 5d pop %rbp |
| 263 |
ffffffff81044222: 48 98 cltq |
| 264 |
ffffffff81044224: c3 retq |
| 265 |
ffffffff81044225: 48 c7 c7 13 53 98 81 mov $0xffffffff81985313,%rdi |
| 266 |
ffffffff8104422c: 31 c0 xor %eax,%eax |
| 267 |
ffffffff8104422e: e8 60 0f 6d 00 callq ffffffff81715193 <printk> |
| 268 |
ffffffff81044233: eb c9 jmp ffffffff810441fe <sys_getppid+0xe> |
| 269 |
ffffffff81044235: 66 66 2e 0f 1f 84 00 data32 nopw %cs:0x0(%rax,%rax,1) |
| 270 |
ffffffff8104423c: 00 00 00 00 |
| 271 |
|
| 272 |
Thus, the disable jump label case adds a 'mov', 'test' and 'jne' instruction |
| 273 |
vs. the jump label case just has a 'no-op' or 'jmp 0'. (The jmp 0, is patched |
| 274 |
to a 5 byte atomic no-op instruction at boot-time.) Thus, the disabled jump |
| 275 |
label case adds:: |
| 276 |
|
| 277 |
6 (mov) + 2 (test) + 2 (jne) = 10 - 5 (5 byte jump 0) = 5 addition bytes. |
| 278 |
|
| 279 |
If we then include the padding bytes, the jump label code saves, 16 total bytes |
| 280 |
of instruction memory for this small function. In this case the non-jump label |
| 281 |
function is 80 bytes long. Thus, we have saved 20% of the instruction |
| 282 |
footprint. We can in fact improve this even further, since the 5-byte no-op |
| 283 |
really can be a 2-byte no-op since we can reach the branch with a 2-byte jmp. |
| 284 |
However, we have not yet implemented optimal no-op sizes (they are currently |
| 285 |
hard-coded). |
| 286 |
|
| 287 |
Since there are a number of static key API uses in the scheduler paths, |
| 288 |
'pipe-test' (also known as 'perf bench sched pipe') can be used to show the |
| 289 |
performance improvement. Testing done on 3.3.0-rc2: |
| 290 |
|
| 291 |
jump label disabled:: |
| 292 |
|
| 293 |
Performance counter stats for 'bash -c /tmp/pipe-test' (50 runs): |
| 294 |
|
| 295 |
855.700314 task-clock # 0.534 CPUs utilized ( +- 0.11% ) |
| 296 |
200,003 context-switches # 0.234 M/sec ( +- 0.00% ) |
| 297 |
0 CPU-migrations # 0.000 M/sec ( +- 39.58% ) |
| 298 |
487 page-faults # 0.001 M/sec ( +- 0.02% ) |
| 299 |
1,474,374,262 cycles # 1.723 GHz ( +- 0.17% ) |
| 300 |
<not supported> stalled-cycles-frontend |
| 301 |
<not supported> stalled-cycles-backend |
| 302 |
1,178,049,567 instructions # 0.80 insns per cycle ( +- 0.06% ) |
| 303 |
208,368,926 branches # 243.507 M/sec ( +- 0.06% ) |
| 304 |
5,569,188 branch-misses # 2.67% of all branches ( +- 0.54% ) |
| 305 |
|
| 306 |
1.601607384 seconds time elapsed ( +- 0.07% ) |
| 307 |
|
| 308 |
jump label enabled:: |
| 309 |
|
| 310 |
Performance counter stats for 'bash -c /tmp/pipe-test' (50 runs): |
| 311 |
|
| 312 |
841.043185 task-clock # 0.533 CPUs utilized ( +- 0.12% ) |
| 313 |
200,004 context-switches # 0.238 M/sec ( +- 0.00% ) |
| 314 |
0 CPU-migrations # 0.000 M/sec ( +- 40.87% ) |
| 315 |
487 page-faults # 0.001 M/sec ( +- 0.05% ) |
| 316 |
1,432,559,428 cycles # 1.703 GHz ( +- 0.18% ) |
| 317 |
<not supported> stalled-cycles-frontend |
| 318 |
<not supported> stalled-cycles-backend |
| 319 |
1,175,363,994 instructions # 0.82 insns per cycle ( +- 0.04% ) |
| 320 |
206,859,359 branches # 245.956 M/sec ( +- 0.04% ) |
| 321 |
4,884,119 branch-misses # 2.36% of all branches ( +- 0.85% ) |
| 322 |
|
| 323 |
1.579384366 seconds time elapsed |
| 324 |
|
| 325 |
The percentage of saved branches is .7%, and we've saved 12% on |
| 326 |
'branch-misses'. This is where we would expect to get the most savings, since |
| 327 |
this optimization is about reducing the number of branches. In addition, we've |
| 328 |
saved .2% on instructions, and 2.8% on cycles and 1.4% on elapsed time. |