[PATCH] x86, MCE, AMD: move invariant code out from loop body

[PATCH] x86, MCE, AMD: move invariant code out from loop body

[Chen Yucong]

On Thu, 2014-10-02 at 16:38 +0200, Borislav Petkov wrote:
> 
> On Mon, Sep 22, 2014 at 09:11:00PM +0200, Borislav Petkov wrote:
> > On Mon, Sep 22, 2014 at 05:23:32PM +0800, Chen Yucong wrote:
> > >  Hi Boris,
> > > 
> > > I have found the following code snippet in mce_amd.c. 
> > > 
> > > /* cpu init entry point, called from mce.c with preempt off */
> > > void mce_amd_feature_init(struct cpuinfo_x86 *c)
> > > {
> > >         ... ...
> > >         for (bank = 0; bank < mca_cfg.banks; ++bank) {
> > >                 for (block = 0; block < NR_BLOCKS; ++block) {
> > >                         ... ...
> > >                         mce_threshold_block_init(&b, offset);
> > >                         mce_threshold_vector = amd_threshold_interrupt;
> > >                 }
> > >         }
> > > }
> > > 
> > > Why should "mce_threshold_vector = amd_threshold_interrupt" be placed in
> > > the inner loop body?
> > 
> > Yeah, it was added sloppily with b276268631af3, I'm not surprised. Feel
> > free to send a fix.
> 
> do you still want to send a fix or should I fix it up quickly?
> 

From: Chen Yucong <slaoub@gmail.com>
Subject: [PATCH] x86, MCE, AMD: move invariant code out from loop body

"mce_threshold_vector = amd_threshold_interrupt;" is loop invariant code
in mce_amd_feature_init(). So it should be moved out from loop body.

Signed-off-by: Chen Yucong <slaoub@gmail.com>
---
 arch/x86/kernel/cpu/mcheck/mce_amd.c |    3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/arch/x86/kernel/cpu/mcheck/mce_amd.c b/arch/x86/kernel/cpu/mcheck/mce_amd.c
index 5d4999f..f727701 100644
--- a/arch/x86/kernel/cpu/mcheck/mce_amd.c
+++ b/arch/x86/kernel/cpu/mcheck/mce_amd.c
@@ -253,9 +253,10 @@ void mce_amd_feature_init(struct cpuinfo_x86 *c)
 			}
 
 			mce_threshold_block_init(&b, offset);
-			mce_threshold_vector = amd_threshold_interrupt;
 		}
 	}
+
+	mce_threshold_vector = amd_threshold_interrupt;
 }
 
 /*
-- 
1.7.10.4

Re: [PATCH] x86, MCE, AMD: move invariant code out from loop body

[Borislav Petkov]

On Thu, Oct 02, 2014 at 11:20:12PM +0800, Chen Yucong wrote:
> From: Chen Yucong <slaoub@gmail.com>
> Subject: [PATCH] x86, MCE, AMD: move invariant code out from loop body
> 
> "mce_threshold_vector = amd_threshold_interrupt;" is loop invariant code
> in mce_amd_feature_init(). So it should be moved out from loop body.
> 
> Signed-off-by: Chen Yucong <slaoub@gmail.com>
> ---
>  arch/x86/kernel/cpu/mcheck/mce_amd.c |    3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/arch/x86/kernel/cpu/mcheck/mce_amd.c b/arch/x86/kernel/cpu/mcheck/mce_amd.c
> index 5d4999f..f727701 100644
> --- a/arch/x86/kernel/cpu/mcheck/mce_amd.c
> +++ b/arch/x86/kernel/cpu/mcheck/mce_amd.c
> @@ -253,9 +253,10 @@ void mce_amd_feature_init(struct cpuinfo_x86 *c)
>  			}
>  
>  			mce_threshold_block_init(&b, offset);
> -			mce_threshold_vector = amd_threshold_interrupt;
>  		}
>  	}
> +
> +	mce_threshold_vector = amd_threshold_interrupt;

Looking at this more, it is theoretically possible that we break out
of the both loops without *any* thresholding registers detected and to
still assign a thresholding interrupt vector which would be clearly
wrong.

Thus I think something like below should be much safer (I tried it with
a label and goto already but it is uglier):

diff --git a/arch/x86/kernel/cpu/mcheck/mce_amd.c b/arch/x86/kernel/cpu/mcheck/mce_amd.c
index 9ce64955559d..9af7bd74828b 100644
--- a/arch/x86/kernel/cpu/mcheck/mce_amd.c
+++ b/arch/x86/kernel/cpu/mcheck/mce_amd.c
@@ -253,7 +253,9 @@ void mce_amd_feature_init(struct cpuinfo_x86 *c)
 			}
 
 			mce_threshold_block_init(&b, offset);
-			mce_threshold_vector = amd_threshold_interrupt;
+
+			if (mce_threshold_vector != amd_threshold_interrupt)
+				mce_threshold_vector = amd_threshold_interrupt;
 		}
 	}
 }

Looking at the asm, we still go and fetch those addresses so not really
a win:

	cmpq	$amd_threshold_interrupt, mce_threshold_vector(%rip)	#, mce_threshold_vector
	je	.L235	#,
	incl	%r13d	# block
	movq	$amd_threshold_interrupt, mce_threshold_vector(%rip)	#, mce_threshold_vector
	cmpl	$9, %r13d	#, block

but this way the code is relatively clean. Unless you can come up with
a nicer, cleaner version to handle the breaking out in the success and
failure case...

Hmmm.

-- 
Regards/Gruss,
    Boris.

Sent from a fat crate under my desk. Formatting is fine.
--

Re: [PATCH] x86, MCE, AMD: move invariant code out from loop body

[Chen Yucong]

On Mon, 2014-10-06 at 23:27 +0200, Borislav Petkov wrote:
> On Thu, Oct 02, 2014 at 11:20:12PM +0800, Chen Yucong wrote:
> > From: Chen Yucong <slaoub@gmail.com>
> > Subject: [PATCH] x86, MCE, AMD: move invariant code out from loop body
> > 
> > "mce_threshold_vector = amd_threshold_interrupt;" is loop invariant code
> > in mce_amd_feature_init(). So it should be moved out from loop body.
> > 
> > Signed-off-by: Chen Yucong <slaoub@gmail.com>
> > ---
> >  arch/x86/kernel/cpu/mcheck/mce_amd.c |    3 ++-
> >  1 file changed, 2 insertions(+), 1 deletion(-)
> > 
> > diff --git a/arch/x86/kernel/cpu/mcheck/mce_amd.c b/arch/x86/kernel/cpu/mcheck/mce_amd.c
> > index 5d4999f..f727701 100644
> > --- a/arch/x86/kernel/cpu/mcheck/mce_amd.c
> > +++ b/arch/x86/kernel/cpu/mcheck/mce_amd.c
> > @@ -253,9 +253,10 @@ void mce_amd_feature_init(struct cpuinfo_x86 *c)
> >  			}
> >  
> >  			mce_threshold_block_init(&b, offset);
> > -			mce_threshold_vector = amd_threshold_interrupt;
> >  		}
> >  	}
> > +
> > +	mce_threshold_vector = amd_threshold_interrupt;
> 
> Looking at this more, it is theoretically possible that we break out
> of the both loops without *any* thresholding registers detected and to
> still assign a thresholding interrupt vector which would be clearly
> wrong.
Yes! In this case, mce_threshold_vector should be `default_threshold_interrupt' rather than
amd_threshold_interrupt.
 
> Thus I think something like below should be much safer (I tried it with
> a label and goto already but it is uglier):
> 
> diff --git a/arch/x86/kernel/cpu/mcheck/mce_amd.c b/arch/x86/kernel/cpu/mcheck/mce_amd.c
> index 9ce64955559d..9af7bd74828b 100644
> --- a/arch/x86/kernel/cpu/mcheck/mce_amd.c
> +++ b/arch/x86/kernel/cpu/mcheck/mce_amd.c
> @@ -253,7 +253,9 @@ void mce_amd_feature_init(struct cpuinfo_x86 *c)
>  			}
>  
>  			mce_threshold_block_init(&b, offset);
> -			mce_threshold_vector = amd_threshold_interrupt;
> +
> +			if (mce_threshold_vector != amd_threshold_interrupt)
> +				mce_threshold_vector = amd_threshold_interrupt;
Perhaps the above assignment operation should be put into 

    if (b.interrupt_capable) {
            ... ...

            if (mce_threshold_vector != amd_threshold_interrupt)
                    mce_threshold_vector = amd_threshold_interrupt;
    }

If IntP (Thresholding Interrupt Supported) bit is zero, this indicates that the reporting
of threshold overflow via interrupt isn't supported. So there's no need to execute the
above assignment operation. 

>  		}
>  	}
>  }
> 
> Looking at the asm, we still go and fetch those addresses so not really
> a win:
> 
> 	cmpq	$amd_threshold_interrupt, mce_threshold_vector(%rip)	#, mce_threshold_vector
> 	je	.L235	#,
> 	incl	%r13d	# block
> 	movq	$amd_threshold_interrupt, mce_threshold_vector(%rip)	#, mce_threshold_vector
> 	cmpl	$9, %r13d	#, block
> 
> but this way the code is relatively clean. Unless you can come up with
> a nicer, cleaner version to handle the breaking out in the success and
> failure case...
Seems like I don't have any better idea than this.

thx!
cyc


From: Chen Yucong <slaoub@gmail.com>
Subject: [PATCH] x86, MCE, AMD: avoid inappropriate assignment operation in
 mce_amd_feature_init

Before executing "mce_threshold_vector = amd_threshold_interrupt;", a few
conditions should be checked for avoiding inappropriate assignment operations,
for example, IntP (Thresholding Interrupt Supported) bit of MCx_MISCi.

Signed-off-by: Chen Yucong <slaoub@gmail.com>
---
 arch/x86/kernel/cpu/mcheck/mce_amd.c |    5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/arch/x86/kernel/cpu/mcheck/mce_amd.c b/arch/x86/kernel/cpu/mcheck/mce_amd.c
index 5d4999f..31bf792 100644
--- a/arch/x86/kernel/cpu/mcheck/mce_amd.c
+++ b/arch/x86/kernel/cpu/mcheck/mce_amd.c
@@ -250,10 +250,13 @@ void mce_amd_feature_init(struct cpuinfo_x86 *c)
 			if (b.interrupt_capable) {
 				int new = (high & MASK_LVTOFF_HI) >> 20;
 				offset  = setup_APIC_mce(offset, new);
+
+				if (offset == new &&
+				     mce_threshold_vector != amd_threshold_interrupt)
+					mce_threshold_vector = amd_threshold_interrupt;
 			}
 
 			mce_threshold_block_init(&b, offset);
-			mce_threshold_vector = amd_threshold_interrupt;
 		}
 	}
 }
-- 
1.7.10.4