[PATCH v2] seccomp: Improve performace by optimizing rmb()

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



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




[Index of Archives]     [Linux Samsung SoC]     [Linux Rockchip SoC]     [Linux Actions SoC]     [Linux for Synopsys ARC Processors]     [Linux NFS]     [Linux NILFS]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]


  Powered by Linux