Static Keys
-----------
By: Jason Baron <jbaron@redhat.com>
0) Abstract
Static keys allows the inclusion of seldom used features in
performance-sensitive fast-path kernel code, via a GCC feature and a code
patching technique. A quick example:
struct static_key key = STATIC_KEY_INIT_FALSE;
...
if (static_key_false(&key))
do unlikely code
else
do likely code
...
static_key_slow_inc();
...
static_key_slow_inc();
...
The static_key_false() branch will be generated into the code with as little
impact to the likely code path as possible.
1) Motivation
Currently, tracepoints are implemented using a conditional branch. The
conditional check requires checking a global variable for each tracepoint.
Although the overhead of this check is small, it increases when the memory
cache comes under pressure (memory cache lines for these global variables may
be shared with other memory accesses). As we increase the number of tracepoints
in the kernel this overhead may become more of an issue. In addition,
tracepoints are often dormant (disabled) and provide no direct kernel
functionality. Thus, it is highly desirable to reduce their impact as much as
possible. Although tracepoints are the original motivation for this work, other
kernel code paths should be able to make use of the static keys facility.
2) Solution
gcc (v4.5) adds a new 'asm goto' statement that allows branching to a label:
http://gcc.gnu.org/ml/gcc-patches/2009-07/msg01556.html
Using the 'asm goto', we can create branches that are either taken or not taken
by default, without the need to check memory. Then, at run-time, we can patch
the branch site to change the branch direction.
For example, if we have a simple branch that is disabled by default:
if (static_key_false(&key))
printk("I am the true branch\n");
Thus, by default the 'printk' will not be emitted. And the code generated will
consist of a single atomic 'no-op' instruction (5 bytes on x86), in the
straight-line code path. When the branch is 'flipped', we will patch the
'no-op' in the straight-line codepath with a 'jump' instruction to the
out-of-line true branch. Thus, changing branch direction is expensive but
branch selection is basically 'free'. That is the basic tradeoff of this
optimization.
This lowlevel patching mechanism is called 'jump label patching', and it gives
the basis for the static keys facility.
3) Static key label API, usage and examples:
In order to make use of this optimization you must first define a key:
struct static_key key;
Which is initialized as:
struct static_key key = STATIC_KEY_INIT_TRUE;
or:
struct static_key key = STATIC_KEY_INIT_FALSE;
If the key is not initialized, it is default false. The 'struct static_key',
must be a 'global'. That is, it can't be allocated on the stack or dynamically
allocated at run-time.
The key is then used in code as:
if (static_key_false(&key))
do unlikely code
else
do likely code
Or:
if (static_key_true(&key))
do likely code
else
do unlikely code
A key that is initialized via 'STATIC_KEY_INIT_FALSE', must be used in a
'static_key_false()' construct. Likewise, a key initialized via
'STATIC_KEY_INIT_TRUE' must be used in a 'static_key_true()' construct. A
single key can be used in many branches, but all the branches must match the
way that the key has been initialized.
The branch(es) can then be switched via:
static_key_slow_inc(&key);
...
static_key_slow_dec(&key);
Thus, 'static_key_slow_inc()' means 'make the branch true', and
'static_key_slow_dec()' means 'make the branch false' with appropriate
reference counting. For example, if the key is initialized true, a
static_key_slow_dec(), will switch the branch to false. And a subsequent
static_key_slow_inc(), will change the branch back to true. Likewise, if the
key is initialized false, a 'static_key_slow_inc()', will change the branch to
true. And then a 'static_key_slow_dec()', will again make the branch false.
An example usage in the kernel is the implementation of tracepoints:
static inline void trace_##name(proto) \
{ \
if (static_key_false(&__tracepoint_##name.key)) \
__DO_TRACE(&__tracepoint_##name, \
TP_PROTO(data_proto), \
TP_ARGS(data_args), \
TP_CONDITION(cond)); \
}
Tracepoints are disabled by default, and can be placed in performance critical
pieces of the kernel. Thus, by using a static key, the tracepoints can have
absolutely minimal impact when not in use.
4) Architecture level code patching interface, 'jump labels'
There are a few functions and macros that architectures must implement in order
to take advantage of this optimization. If there is no architecture support, we
simply fall back to a traditional, load, test, and jump sequence.
* select HAVE_ARCH_JUMP_LABEL, see: arch/x86/Kconfig
* #define JUMP_LABEL_NOP_SIZE, see: arch/x86/include/asm/jump_label.h
* __always_inline bool arch_static_branch(struct static_key *key), see:
arch/x86/include/asm/jump_label.h
* void arch_jump_label_transform(struct jump_entry *entry, enum jump_label_type type),
see: arch/x86/kernel/jump_label.c
* __init_or_module void arch_jump_label_transform_static(struct jump_entry *entry, enum jump_label_type type),
see: arch/x86/kernel/jump_label.c
* struct jump_entry, see: arch/x86/include/asm/jump_label.h
5) Static keys / jump label analysis, results (x86_64):
As an example, let's add the following branch to 'getppid()', such that the
system call now looks like:
SYSCALL_DEFINE0(getppid)
{
int pid;
+ if (static_key_false(&key))
+ printk("I am the true branch\n");
rcu_read_lock();
pid = task_tgid_vnr(rcu_dereference(current->real_parent));
rcu_read_unlock();
return pid;
}
The resulting instructions with jump labels generated by GCC is:
ffffffff81044290 <sys_getppid>:
ffffffff81044290: 55 push %rbp
ffffffff81044291: 48 89 e5 mov %rsp,%rbp
ffffffff81044294: e9 00 00 00 00 jmpq ffffffff81044299 <sys_getppid+0x9>
ffffffff81044299: 65 48 8b 04 25 c0 b6 mov %gs:0xb6c0,%rax
ffffffff810442a0: 00 00
ffffffff810442a2: 48 8b 80 80 02 00 00 mov 0x280(%rax),%rax
ffffffff810442a9: 48 8b 80 b0 02 00 00 mov 0x2b0(%rax),%rax
ffffffff810442b0: 48 8b b8 e8 02 00 00 mov 0x2e8(%rax),%rdi
ffffffff810442b7: e8 f4 d9 00 00 callq ffffffff81051cb0 <pid_vnr>
ffffffff810442bc: 5d pop %rbp
ffffffff810442bd: 48 98 cltq
ffffffff810442bf: c3 retq
ffffffff810442c0: 48 c7 c7 e3 54 98 81 mov $0xffffffff819854e3,%rdi
ffffffff810442c7: 31 c0 xor %eax,%eax
ffffffff810442c9: e8 71 13 6d 00 callq ffffffff8171563f <printk>
ffffffff810442ce: eb c9 jmp ffffffff81044299 <sys_getppid+0x9>
Without the jump label optimization it looks like:
ffffffff810441f0 <sys_getppid>:
ffffffff810441f0: 8b 05 8a 52 d8 00 mov 0xd8528a(%rip),%eax # ffffffff81dc9480 <key>
ffffffff810441f6: 55 push %rbp
ffffffff810441f7: 48 89 e5 mov %rsp,%rbp
ffffffff810441fa: 85 c0 test %eax,%eax
ffffffff810441fc: 75 27 jne ffffffff81044225 <sys_getppid+0x35>
ffffffff810441fe: 65 48 8b 04 25 c0 b6 mov %gs:0xb6c0,%rax
ffffffff81044205: 00 00
ffffffff81044207: 48 8b 80 80 02 00 00 mov 0x280(%rax),%rax
ffffffff8104420e: 48 8b 80 b0 02 00 00 mov 0x2b0(%rax),%rax
ffffffff81044215: 48 8b b8 e8 02 00 00 mov 0x2e8(%rax),%rdi
ffffffff8104421c: e8 2f da 00 00 callq ffffffff81051c50 <pid_vnr>
ffffffff81044221: 5d pop %rbp
ffffffff81044222: 48 98 cltq
ffffffff81044224: c3 retq
ffffffff81044225: 48 c7 c7 13 53 98 81 mov $0xffffffff81985313,%rdi
ffffffff8104422c: 31 c0 xor %eax,%eax
ffffffff8104422e: e8 60 0f 6d 00 callq ffffffff81715193 <printk>
ffffffff81044233: eb c9 jmp ffffffff810441fe <sys_getppid+0xe>
ffffffff81044235: 66 66 2e 0f 1f 84 00 data32 nopw %cs:0x0(%rax,%rax,1)
ffffffff8104423c: 00 00 00 00
Thus, the disable jump label case adds a 'mov', 'test' and 'jne' instruction
vs. the jump label case just has a 'no-op' or 'jmp 0'. (The jmp 0, is patched
to a 5 byte atomic no-op instruction at boot-time.) Thus, the disabled jump
label case adds:
6 (mov) + 2 (test) + 2 (jne) = 10 - 5 (5 byte jump 0) = 5 addition bytes.
If we then include the padding bytes, the jump label code saves, 16 total bytes
of instruction memory for this small function. In this case the non-jump label
function is 80 bytes long. Thus, we have saved 20% of the instruction
footprint. We can in fact improve this even further, since the 5-byte no-op
really can be a 2-byte no-op since we can reach the branch with a 2-byte jmp.
However, we have not yet implemented optimal no-op sizes (they are currently
hard-coded).
Since there are a number of static key API uses in the scheduler paths,
'pipe-test' (also known as 'perf bench sched pipe') can be used to show the
performance improvement. Testing done on 3.3.0-rc2:
jump label disabled:
Performance counter stats for 'bash -c /tmp/pipe-test' (50 runs):
855.700314 task-clock # 0.534 CPUs utilized ( +- 0.11% )
200,003 context-switches # 0.234 M/sec ( +- 0.00% )
0 CPU-migrations # 0.000 M/sec ( +- 39.58% )
487 page-faults # 0.001 M/sec ( +- 0.02% )
1,474,374,262 cycles # 1.723 GHz ( +- 0.17% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
1,178,049,567 instructions # 0.80 insns per cycle ( +- 0.06% )
208,368,926 branches # 243.507 M/sec ( +- 0.06% )
5,569,188 branch-misses # 2.67% of all branches ( +- 0.54% )
1.601607384 seconds time elapsed ( +- 0.07% )
jump label enabled:
Performance counter stats for 'bash -c /tmp/pipe-test' (50 runs):
841.043185 task-clock # 0.533 CPUs utilized ( +- 0.12% )
200,004 context-switches # 0.238 M/sec ( +- 0.00% )
0 CPU-migrations # 0.000 M/sec ( +- 40.87% )
487 page-faults # 0.001 M/sec ( +- 0.05% )
1,432,559,428 cycles # 1.703 GHz ( +- 0.18% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
1,175,363,994 instructions # 0.82 insns per cycle ( +- 0.04% )
206,859,359 branches # 245.956 M/sec ( +- 0.04% )
4,884,119 branch-misses # 2.36% of all branches ( +- 0.85% )
1.579384366 seconds time elapsed
The percentage of saved branches is .7%, and we've saved 12% on
'branch-misses'. This is where we would expect to get the most savings, since
this optimization is about reducing the number of branches. In addition, we've
saved .2% on instructions, and 2.8% on cycles and 1.4% on elapsed time.