According to kees's suggest, we started with the patch that just replaces
rmb() with smp_rmb() and did a performace test with UnixBench. The results
showed the overhead about 2.53% in rmb() test compared to the smp_rmb()
one, in a x86-64 kernel with CONFIG_SMP enabled running inside a qemu-kvm
vm. The test is a "syscall" testcase in UnixBench, which executes 5
syscalls in a loop during a certain timeout (100 second in our test) and
counts the total number of executions of this 5-syscall sequence. We set a
seccomp filter with all allow rule for all used syscalls in this test
(which will go bitmap path) to make sure the rmb() will be executed. The
details for the test:
with rmb():
/txm # ./syscall_allow_min 100
COUNT|35861159|1|lps
/txm # ./syscall_allow_min 100
COUNT|35545501|1|lps
/txm # ./syscall_allow_min 100
COUNT|35664495|1|lps
with smp_rmb():
/txm # ./syscall_allow_min 100
COUNT|36552771|1|lps
/txm # ./syscall_allow_min 100
COUNT|36491247|1|lps
/txm # ./syscall_allow_min 100
COUNT|36504746|1|lps
For a x86-64 kernel with CONFIG_SMP enabled, the smp_rmb() is just a
compiler barrier() which have no impact in runtime, while rmb() is a
lfence which will prevent all memory access operations (not just load
according the recently claim by Intel) behind itself. We can also figure
it out in disassembly:
with rmb():
0000000000001430 <__seccomp_filter>:
1430: 41 57 push %r15
1432: 41 56 push %r14
1434: 41 55 push %r13
1436: 41 54 push %r12
1438: 55 push %rbp
1439: 53 push %rbx
143a: 48 81 ec 90 00 00 00 sub $0x90,%rsp
1441: 89 7c 24 10 mov %edi,0x10(%rsp)
1445: 89 54 24 14 mov %edx,0x14(%rsp)
1449: 65 48 8b 04 25 28 00 mov %gs:0x28,%rax
1450: 00 00
1452: 48 89 84 24 88 00 00 mov %rax,0x88(%rsp)
1459: 00
145a: 31 c0 xor %eax,%eax
* 145c: 0f ae e8 lfence
145f: 48 85 f6 test %rsi,%rsi
1462: 49 89 f4 mov %rsi,%r12
1465: 0f 84 42 03 00 00 je 17ad <__seccomp_filter+0x37d>
146b: 65 48 8b 04 25 00 00 mov %gs:0x0,%rax
1472: 00 00
1474: 48 8b 98 80 07 00 00 mov 0x780(%rax),%rbx
147b: 48 85 db test %rbx,%rbx
with smp_rmb();
0000000000001430 <__seccomp_filter>:
1430: 41 57 push %r15
1432: 41 56 push %r14
1434: 41 55 push %r13
1436: 41 54 push %r12
1438: 55 push %rbp
1439: 53 push %rbx
143a: 48 81 ec 90 00 00 00 sub $0x90,%rsp
1441: 89 7c 24 10 mov %edi,0x10(%rsp)
1445: 89 54 24 14 mov %edx,0x14(%rsp)
1449: 65 48 8b 04 25 28 00 mov %gs:0x28,%rax
1450: 00 00
1452: 48 89 84 24 88 00 00 mov %rax,0x88(%rsp)
1459: 00
145a: 31 c0 xor %eax,%eax
145c: 48 85 f6 test %rsi,%rsi
145f: 49 89 f4 mov %rsi,%r12
1462: 0f 84 42 03 00 00 je 17aa <__seccomp_filter+0x37a>
1468: 65 48 8b 04 25 00 00 mov %gs:0x0,%rax
146f: 00 00
1471: 48 8b 98 80 07 00 00 mov 0x780(%rax),%rbx
1478: 48 85 db test %rbx,%rbx
We will go further for the next optimize patch, if you guys thinks this
smp_rmb() refactor is appropriate.
v1 -> v2:
- only replace rmb() with smp_rmb()
- provide the performance test number
RFC -> v1:
- replace rmb() with smp_rmb()
- move the smp_rmb() logic to the middle between TIF_SECCOMP and mode
Signed-off-by: wanghongzhe <wanghongzhe@xxxxxxxxxx>
---
kernel/seccomp.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/kernel/seccomp.c b/kernel/seccomp.c
index 952dc1c90229..8505b438a590 100644
--- a/kernel/seccomp.c
+++ b/kernel/seccomp.c
@@ -1164,7 +1164,7 @@ static int __seccomp_filter(int this_syscall, const struct seccomp_data *sd,
* Make sure that any changes to mode from another thread have
* been seen after SYSCALL_WORK_SECCOMP was seen.
*/
- rmb();
+ smp_rmb();
if (!sd) {
populate_seccomp_data(&sd_local);
--
2.19.1
