powerpc: Improve scheduling of system call entry instructions
authorAnton Blanchard <anton@samba.org>
Sun, 8 May 2011 21:36:44 +0000 (21:36 +0000)
committerBenjamin Herrenschmidt <benh@kernel.crashing.org>
Thu, 19 May 2011 04:30:42 +0000 (14:30 +1000)
After looking at our system call path, Mary Brown suggested that we
should put all mfspr SRR* instructions before any mtspr SRR*.

To test this I used a very simple null syscall (actually getppid)
testcase at http://ozlabs.org/~anton/junkcode/null_syscall.c

I tested with the following changes against the pseries_defconfig:

CONFIG_VIRT_CPU_ACCOUNTING=n
CONFIG_AUDIT=n

to remove the overhead of virtual CPU accounting and syscall
auditing.

POWER6:
baseline:       mean = 757.2 cycles       sd = 2.108
modified:       mean = 759.1 cycles       sd = 2.020

POWER7:
baseline:       mean = 411.4 cycles       sd = 0.138
modified:       mean = 404.1 cycles       sd = 0.109

So we have 1.77% improvement on POWER7 which looks significant. The
POWER6 suggest a 0.25% slowdown, but the results are within 1
standard deviation and may be in the noise.

Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
arch/powerpc/kernel/exceptions-64s.S

index 0ec3b42717d7ea2694a069b1fa8162023e2ec136..a85f4874cba771fca11887ce4be593bde2c5728a 100644 (file)
@@ -211,11 +211,11 @@ END_FTR_SECTION_IFSET(CPU_FTR_REAL_LE)
        mr      r9,r13
        GET_PACA(r13)
        mfspr   r11,SPRN_SRR0
-       ld      r12,PACAKBASE(r13)
-       ld      r10,PACAKMSR(r13)
-       LOAD_HANDLER(r12, system_call_entry)
-       mtspr   SPRN_SRR0,r12
        mfspr   r12,SPRN_SRR1
+       ld      r10,PACAKBASE(r13)
+       LOAD_HANDLER(r10, system_call_entry)
+       mtspr   SPRN_SRR0,r10
+       ld      r10,PACAKMSR(r13)
        mtspr   SPRN_SRR1,r10
        rfid
        b       .       /* prevent speculative execution */