By delaying the setting of MSR_RI, a 1% improvment is optained on
null_syscall selftest on an mpc8321.
Without this patch:
root@vgoippro:~# ./null_syscall
1134.33 ns 378.11 cycles
With this patch:
root@vgoippro:~# ./null_syscall
1121.85 ns 373.95 cycles
The drawback is that a machine check during that period
would be unrecoverable, but as only main memory is accessed
during that period, it shouldn't be a concern.
Signed-off-by: Christophe Leroy <redacted>
---
arch/powerpc/kernel/head_32.S | 2 --
1 file changed, 2 deletions(-)
From: Michael Ellerman <mpe@ellerman.id.au> Date: 2019-02-01 11:10:38
Christophe Leroy [off-list ref] writes:
By delaying the setting of MSR_RI, a 1% improvment is optained on
null_syscall selftest on an mpc8321.
Without this patch:
root@vgoippro:~# ./null_syscall
1134.33 ns 378.11 cycles
With this patch:
root@vgoippro:~# ./null_syscall
1121.85 ns 373.95 cycles
The drawback is that a machine check during that period
would be unrecoverable, but as only main memory is accessed
during that period, it shouldn't be a concern.
On 64-bit server CPUs accessing main memory can cause a UE
(Uncorrectable Error) which can trigger a machine check.
So it may still be a concern, it depends how paranoid you are.
By delaying the setting of MSR_RI, a 1% improvment is optained on
null_syscall selftest on an mpc8321.
Without this patch:
root@vgoippro:~# ./null_syscall
1134.33 ns 378.11 cycles
With this patch:
root@vgoippro:~# ./null_syscall
1121.85 ns 373.95 cycles
The drawback is that a machine check during that period
would be unrecoverable, but as only main memory is accessed
during that period, it shouldn't be a concern.
On 64-bit server CPUs accessing main memory can cause a UE
(Uncorrectable Error) which can trigger a machine check.
So it may still be a concern, it depends how paranoid you are.
Where does RI get enabled? I don't see it anywhere obvious.
MSR_RI is part of MSR_KERNEL, it gets then enabled when reenabling MMU when calling the exception handler.
#define EXC_XFER_TEMPLATE(n, hdlr, trap, copyee, tfer, ret) \
li r10,trap; \
stw r10,_TRAP(r11); \
li r10,MSR_KERNEL; \
copyee(r10, r9); \
bl tfer; \
i##n: \
.long hdlr; \
.long ret
where tfer = transfer_to_handler.
In transfer_to_handler (kernel/entry_32.S) you have:
transfer_to_handler_cont:
3:
mflr r9
lwz r11,0(r9) /* virtual address of handler */
lwz r9,4(r9) /* where to go when done */
[...]
mtspr SPRN_SRR0,r11
mtspr SPRN_SRR1,r10
mtlr r9
SYNC
RFI /* jump to handler, enable MMU */
So MSR_RI is restored above as r10 contains MSR_KERNEL [ | MSR_EE ]
Christophe
By delaying the setting of MSR_RI, a 1% improvment is optained on
null_syscall selftest on an mpc8321.
Without this patch:
root@vgoippro:~# ./null_syscall
1134.33 ns 378.11 cycles
With this patch:
root@vgoippro:~# ./null_syscall
1121.85 ns 373.95 cycles
The drawback is that a machine check during that period
would be unrecoverable, but as only main memory is accessed
during that period, it shouldn't be a concern.
On 64-bit server CPUs accessing main memory can cause a UE
(Uncorrectable Error) which can trigger a machine check.
So it may still be a concern, it depends how paranoid you are.
stw r1,GPR1(r11); \
stw r1,0(r11); \
tovirt(r1,r11); /* set new kernel sp */ \
- li r10,MSR_KERNEL & ~(MSR_IR|MSR_DR); /* can take exceptions */ \
- MTMSRD(r10); /* (except for mach check in rtas) */ \
stw r0,GPR0(r11); \
lis r10,STACK_FRAME_REGS_MARKER@ha; /* exception frame marker */ \
addi r10,r10,STACK_FRAME_REGS_MARKER@l; \
Where does RI get enabled? I don't see it anywhere obvious.
MSR_RI is part of MSR_KERNEL, it gets then enabled when reenabling MMU when calling the exception handler.
#define EXC_XFER_TEMPLATE(n, hdlr, trap, copyee, tfer, ret) \
li r10,trap; \
stw r10,_TRAP(r11); \
li r10,MSR_KERNEL; \
copyee(r10, r9); \
bl tfer; \
i##n: \
.long hdlr; \
.long ret
where tfer = transfer_to_handler.
In transfer_to_handler (kernel/entry_32.S) you have:
transfer_to_handler_cont:
3:
mflr r9
lwz r11,0(r9) /* virtual address of handler */
lwz r9,4(r9) /* where to go when done */
[...]
mtspr SPRN_SRR0,r11
mtspr SPRN_SRR1,r10
mtlr r9
SYNC
RFI /* jump to handler, enable MMU */
So MSR_RI is restored above as r10 contains MSR_KERNEL [ | MSR_EE ]
Looks like fast_exception_return, which is called by hash page handlers at least, expects MSR_RI to be set. Allthough it works well on 603 (because it doesn't hash), I would most likely not work on others.
This 1% improvment is not worth it, I give up for now.