Patchwork powerpc/booke: Eliminate rfi from exception entry path.

login
register
mail settings
Submitter Scott Wood
Date July 11, 2012, 12:34 a.m.
Message ID <20120711003454.GA22757@tyr.buserror.net>
Download mbox | patch
Permalink /patch/170323/
State Under Review
Delegated to: Scott Wood
Headers show

Comments

Scott Wood - July 11, 2012, 12:34 a.m.
Unlike classic, we don't really need the MSR change to be atomic with the
branch.  This eliminates a trap as a KVM guest (in the absence of
hardware hypervisor extensions), where mtmsr is paravirtualized but rfi
is not.  For a virtualized guest without any paravirtualization, this
eliminates an additional two traps (SRR0/1).

Signed-off-by: Scott Wood <scottwood@freescale.com>
---
 arch/powerpc/kernel/entry_32.S |   16 ++++++++++++++++
 1 files changed, 16 insertions(+), 0 deletions(-)
Benjamin Herrenschmidt - July 11, 2012, 12:36 a.m.
On Tue, 2012-07-10 at 19:34 -0500, Scott Wood wrote:
> Unlike classic, we don't really need the MSR change to be atomic with the
> branch.  This eliminates a trap as a KVM guest (in the absence of
> hardware hypervisor extensions), where mtmsr is paravirtualized but rfi
> is not.  For a virtualized guest without any paravirtualization, this
> eliminates an additional two traps (SRR0/1).

In fact, I wonder, what do we write into the MSR at this point that
wasn't already in it in BookE ? RI ? I wonder if we could get away
without the mtmsr alltogether...

Cheers,
Ben.

> Signed-off-by: Scott Wood <scottwood@freescale.com>
> ---
>  arch/powerpc/kernel/entry_32.S |   16 ++++++++++++++++
>  1 files changed, 16 insertions(+), 0 deletions(-)
> 
> diff --git a/arch/powerpc/kernel/entry_32.S b/arch/powerpc/kernel/entry_32.S
> index ba3aeb4..6bb637c 100644
> --- a/arch/powerpc/kernel/entry_32.S
> +++ b/arch/powerpc/kernel/entry_32.S
> @@ -193,6 +193,9 @@ transfer_to_handler_cont:
>  	lwz	r11,0(r9)		/* virtual address of handler */
>  	lwz	r9,4(r9)		/* where to go when done */
>  #ifdef CONFIG_TRACE_IRQFLAGS
> +#ifdef CONFIG_BOOKE
> +	mtmsr	r10
> +#else
>  	lis	r12,reenable_mmu@h
>  	ori	r12,r12,reenable_mmu@l
>  	mtspr	SPRN_SRR0,r12
> @@ -201,6 +204,7 @@ transfer_to_handler_cont:
>  	RFI
>  reenable_mmu:				/* re-enable mmu so we can */
>  	mfmsr	r10
> +#endif /* !CONFIG_BOOKE */
>  	lwz	r12,_MSR(r1)
>  	xor	r10,r10,r12
>  	andi.	r10,r10,MSR_EE		/* Did EE change? */
> @@ -247,11 +251,23 @@ reenable_mmu:				/* re-enable mmu so we can */
>  	mtlr	r9
>  	bctr				/* jump to handler */
>  #else /* CONFIG_TRACE_IRQFLAGS */
> +#ifdef CONFIG_BOOKE
> +	/*
> +	 * We're not changing address space on Book E, and the extra rfi
> +	 * can hurt when virtualized without hardware support -- whereas
> +	 * mtmsr can be paravirtualized.
> +	 */
> +	mtmsr	r10
> +	mtctr	r11
> +	mtlr	r9
> +	bctr
> +#else
>  	mtspr	SPRN_SRR0,r11
>  	mtspr	SPRN_SRR1,r10
>  	mtlr	r9
>  	SYNC
>  	RFI				/* jump to handler, enable MMU */
> +#endif /* !CONFIG_BOOKE */
>  #endif /* CONFIG_TRACE_IRQFLAGS */
>  
>  #if defined (CONFIG_6xx) || defined(CONFIG_E500)
Scott Wood - July 11, 2012, 12:41 a.m.
On 07/10/2012 07:36 PM, Benjamin Herrenschmidt wrote:
> On Tue, 2012-07-10 at 19:34 -0500, Scott Wood wrote:
>> Unlike classic, we don't really need the MSR change to be atomic with the
>> branch.  This eliminates a trap as a KVM guest (in the absence of
>> hardware hypervisor extensions), where mtmsr is paravirtualized but rfi
>> is not.  For a virtualized guest without any paravirtualization, this
>> eliminates an additional two traps (SRR0/1).
> 
> In fact, I wonder, what do we write into the MSR at this point that
> wasn't already in it in BookE ? RI ? I wonder if we could get away
> without the mtmsr alltogether...

Doesn't EE get set there for some exceptions?

-Scott
Alexander Graf - July 11, 2012, 12:44 a.m.
On 11.07.2012, at 02:34, Scott Wood wrote:

> Unlike classic, we don't really need the MSR change to be atomic with the
> branch.  This eliminates a trap as a KVM guest (in the absence of
> hardware hypervisor extensions), where mtmsr is paravirtualized but rfi
> is not.  For a virtualized guest without any paravirtualization, this
> eliminates an additional two traps (SRR0/1).
> 
> Signed-off-by: Scott Wood <scottwood@freescale.com>
> ---
> arch/powerpc/kernel/entry_32.S |   16 ++++++++++++++++
> 1 files changed, 16 insertions(+), 0 deletions(-)
> 
> diff --git a/arch/powerpc/kernel/entry_32.S b/arch/powerpc/kernel/entry_32.S
> index ba3aeb4..6bb637c 100644
> --- a/arch/powerpc/kernel/entry_32.S
> +++ b/arch/powerpc/kernel/entry_32.S
> @@ -193,6 +193,9 @@ transfer_to_handler_cont:
> 	lwz	r11,0(r9)		/* virtual address of handler */
> 	lwz	r9,4(r9)		/* where to go when done */
> #ifdef CONFIG_TRACE_IRQFLAGS
> +#ifdef CONFIG_BOOKE
> +	mtmsr	r10
> +#else
> 	lis	r12,reenable_mmu@h
> 	ori	r12,r12,reenable_mmu@l
> 	mtspr	SPRN_SRR0,r12
> @@ -201,6 +204,7 @@ transfer_to_handler_cont:
> 	RFI
> reenable_mmu:				/* re-enable mmu so we can */
> 	mfmsr	r10
> +#endif /* !CONFIG_BOOKE */
> 	lwz	r12,_MSR(r1)
> 	xor	r10,r10,r12
> 	andi.	r10,r10,MSR_EE		/* Did EE change? */
> @@ -247,11 +251,23 @@ reenable_mmu:				/* re-enable mmu so we can */
> 	mtlr	r9
> 	bctr				/* jump to handler */
> #else /* CONFIG_TRACE_IRQFLAGS */
> +#ifdef CONFIG_BOOKE
> +	/*
> +	 * We're not changing address space on Book E, and the extra rfi
> +	 * can hurt when virtualized without hardware support -- whereas
> +	 * mtmsr can be paravirtualized.

We can always paravirtualize RFI as well if it makes sense.


Alex
Scott Wood - July 11, 2012, 12:47 a.m.
On 07/10/2012 07:44 PM, Alexander Graf wrote:
> 
> On 11.07.2012, at 02:34, Scott Wood wrote:
>> +#ifdef CONFIG_BOOKE
>> +	/*
>> +	 * We're not changing address space on Book E, and the extra rfi
>> +	 * can hurt when virtualized without hardware support -- whereas
>> +	 * mtmsr can be paravirtualized.
> 
> We can always paravirtualize RFI as well if it makes sense.

I'm not sure that's possible.  We thought about it a while back, but
IIRC the difficulty was not leaving a register clobbered.

-Scott
Benjamin Herrenschmidt - July 11, 2012, 12:53 a.m.
On Tue, 2012-07-10 at 19:41 -0500, Scott Wood wrote:
> On 07/10/2012 07:36 PM, Benjamin Herrenschmidt wrote:
> > On Tue, 2012-07-10 at 19:34 -0500, Scott Wood wrote:
> >> Unlike classic, we don't really need the MSR change to be atomic with the
> >> branch.  This eliminates a trap as a KVM guest (in the absence of
> >> hardware hypervisor extensions), where mtmsr is paravirtualized but rfi
> >> is not.  For a virtualized guest without any paravirtualization, this
> >> eliminates an additional two traps (SRR0/1).
> > 
> > In fact, I wonder, what do we write into the MSR at this point that
> > wasn't already in it in BookE ? RI ? I wonder if we could get away
> > without the mtmsr alltogether...
> 
> Doesn't EE get set there for some exceptions?

It does, tho arguably it shouldn't in most cases :-) I'm happy to turn a
bunch of these into explicit local_irq_enable() in the C code though
which will turn into a wrteei which is more efficient on BookE.

Cheers,
Ben.
Benjamin Herrenschmidt - July 11, 2012, 12:54 a.m.
On Tue, 2012-07-10 at 19:47 -0500, Scott Wood wrote:
> On 07/10/2012 07:44 PM, Alexander Graf wrote:
> > 
> > On 11.07.2012, at 02:34, Scott Wood wrote:
> >> +#ifdef CONFIG_BOOKE
> >> +	/*
> >> +	 * We're not changing address space on Book E, and the extra rfi
> >> +	 * can hurt when virtualized without hardware support -- whereas
> >> +	 * mtmsr can be paravirtualized.
> > 
> > We can always paravirtualize RFI as well if it makes sense.
> 
> I'm not sure that's possible.  We thought about it a while back, but
> IIRC the difficulty was not leaving a register clobbered.

Besides mtmsr is slow on real HW as well. Also paravirt as done today
for complex instructions like mtmsr is racy :-) (I already had a chat
about that with Alex a while back, we might want to re-consider what
kind of fix can be done at some point).

Cheers,
Ben.

Patch

diff --git a/arch/powerpc/kernel/entry_32.S b/arch/powerpc/kernel/entry_32.S
index ba3aeb4..6bb637c 100644
--- a/arch/powerpc/kernel/entry_32.S
+++ b/arch/powerpc/kernel/entry_32.S
@@ -193,6 +193,9 @@  transfer_to_handler_cont:
 	lwz	r11,0(r9)		/* virtual address of handler */
 	lwz	r9,4(r9)		/* where to go when done */
 #ifdef CONFIG_TRACE_IRQFLAGS
+#ifdef CONFIG_BOOKE
+	mtmsr	r10
+#else
 	lis	r12,reenable_mmu@h
 	ori	r12,r12,reenable_mmu@l
 	mtspr	SPRN_SRR0,r12
@@ -201,6 +204,7 @@  transfer_to_handler_cont:
 	RFI
 reenable_mmu:				/* re-enable mmu so we can */
 	mfmsr	r10
+#endif /* !CONFIG_BOOKE */
 	lwz	r12,_MSR(r1)
 	xor	r10,r10,r12
 	andi.	r10,r10,MSR_EE		/* Did EE change? */
@@ -247,11 +251,23 @@  reenable_mmu:				/* re-enable mmu so we can */
 	mtlr	r9
 	bctr				/* jump to handler */
 #else /* CONFIG_TRACE_IRQFLAGS */
+#ifdef CONFIG_BOOKE
+	/*
+	 * We're not changing address space on Book E, and the extra rfi
+	 * can hurt when virtualized without hardware support -- whereas
+	 * mtmsr can be paravirtualized.
+	 */
+	mtmsr	r10
+	mtctr	r11
+	mtlr	r9
+	bctr
+#else
 	mtspr	SPRN_SRR0,r11
 	mtspr	SPRN_SRR1,r10
 	mtlr	r9
 	SYNC
 	RFI				/* jump to handler, enable MMU */
+#endif /* !CONFIG_BOOKE */
 #endif /* CONFIG_TRACE_IRQFLAGS */
 
 #if defined (CONFIG_6xx) || defined(CONFIG_E500)