i386: MCE updates

PATCH In review Joerg Roedel 4 participants 15 messages Oct 9, 2007 – Oct 11, 2007

lore

Share & tools

lore

Joerg Roedel

author

2007-10-09 12:52 UTC #0 lore raw

cover i386: MCE updates

This patchset includes two patches.

1. Checks for the MCA fetures as early as possible.
2. Coding style cleanup.

Joerg Roedel

author

2007-10-09 12:50 UTC #1 lore raw

PATCH 1/2 i386: mce cleanup part1: functional change

MCG_CAP never reports a negative count of available error-reporting banks. Therefore, make it unsigned.
Check for MCA/MCE feature bits as early as possible.

Signed-off-by Christoph Egger <[email protected]>

Signed-off-by Joerg Roedel <[email protected]>

---
 arch/i386/kernel/cpu/mcheck/mce.c |   13 ++++++++++++-
 arch/i386/kernel/cpu/mcheck/mce.h |    2 +-
 2 files changed, 13 insertions(+), 2 deletions(-)

arch/i386/kernel/cpu/mcheck/mce.c +12 -1

index 34c781e..bad2c64 100644
@@ -17,7 +17,7 @@
 #include "mce.h"
 
 int mce_disabled = 0;
-int nr_mce_banks;
+unsigned int nr_mce_banks;
 
 EXPORT_SYMBOL_GPL(nr_mce_banks);	/* non-fatal.o */

@@ -33,9 +33,20 @@ void fastcall (*machine_check_vector)(struct pt_regs *, long error_code) = unexp
 /* This has to be run for each processor */
 void mcheck_init(struct cpuinfo_x86 *c)
 {
+	uint32_t mca, mce;
+
 	if (mce_disabled==1)
 		return;
 
+	mca = cpu_has(c, X86_FEATURE_MCA);
+	mce = cpu_has(c, X86_FEATURE_MCE);
+
+	if (!mca || !mce) {
+		printk(KERN_INFO "CPU%i: No machine check support available\n",
+			smp_processor_id());
+		return;
+	}
+
 	switch (c->x86_vendor) {
 		case X86_VENDOR_AMD:
 			amd_mcheck_init(c);

arch/i386/kernel/cpu/mcheck/mce.h +1 -1

index 81fb6e2..9cbe812 100644
@@ -10,5 +10,5 @@ void winchip_mcheck_init(struct cpuinfo_x86 *c);
 /* Call the installed machine check handler for this CPU setup. */
 extern fastcall void (*machine_check_vector)(struct pt_regs *, long error_code);
 
-extern int nr_mce_banks;
+extern unsigned int nr_mce_banks;

Signature · 1 line

1.5.2.5

Joerg Roedel

author

2007-10-09 16:06 UTC #3 lore raw

Re: [PATCH 1/2] i386: mce cleanup part1: functional change

Quoted 25 lines · Oleg Verych

* Tue, 9 Oct 2007 14:49:55 +0200
[]
Quoted 8 lines
arch/i386/kernel/cpu/mcheck/mce.c +2 -0
@@ -33,9 +33,20 @@ void fastcall (*machine_check_vector)(struct pt_regs *, long error_code) = unexp
 /* This has to be run for each processor */
 void mcheck_init(struct cpuinfo_x86 *c)
 {
+	uint32_t mca, mce;
+
 	if (mce_disabled==1)
 		return;
arch/i386/kernel/cpu/mcheck/mce.c void fastcall (*machine_check_vector)(struct pt_regs *, long error_code) = unexp · line 49
Quoted 10 lines
arch/i386/kernel/cpu/mcheck/mce.c +9 -0
 
+	mca = cpu_has(c, X86_FEATURE_MCA);
+	mce = cpu_has(c, X86_FEATURE_MCE);
+
+	if (!mca || !mce) {
+		printk(KERN_INFO "CPU%i: No machine check support available\n",
+			smp_processor_id());
+		return;
+	}
+
cpu_has() returns int,
but would it be better to have something like

Quoted 4 lines

if (!mce_disabled &&

	    !(c->x86_capability & (X86_FEATURE_MCA | X86_FEATURE_MCE)) {
		printk(KERN_INFO "CPU%i: No machine check support available\n",
			smp_processor_id());

This looks complicated and is harder to read. Its exactly the purpose of the cpu_has() macro to avoid such constructs.

Quoted 3 lines

		return;
	} else
		return;

Return unconditionaly here?

Signature · 6 lines

           |           AMD Saxony Limited Liability Company & Co. KG
 Operating |         Wilschdorfer Landstr. 101, 01109 Dresden, Germany
 System    |                  Register Court Dresden: HRA 4896
 Research  |              General Partner authorized to represent:
 Center    |             AMD Saxony LLC (Wilmington, Delaware, US)
           | General Manager of AMD Saxony LLC: Dr. Hans-R. Deppe, Thomas McCoy

Oleg Verych

2007-10-09 16:18 UTC #4 lore raw

Re: [PATCH 1/2] i386: mce cleanup part1: functional change

Quoted 10 lines · Joerg Roedel

Quoted 7 lines
cpu_has() returns int,
but would it be better to have something like
if (!mce_disabled &&
	    !(c->x86_capability & (X86_FEATURE_MCA | X86_FEATURE_MCE)) {
		printk(KERN_INFO "CPU%i: No machine check support available\n",
			smp_processor_id());
This looks complicated and is harder to read. Its exactly the purpose of the cpu_has() macro to avoid such constructs.

It is done via test_bit(), which is designed for IO access with all that `const volatile' stuff, 2 x unnecessary, can't be optimized here (IMHO).
____

Joerg Roedel

author

2007-10-09 17:28 UTC #5 lore raw

Re: [PATCH 1/2] i386: mce cleanup part1: functional change

Quoted 9 lines · Oleg Verych

Quoted 8 lines · Joerg Roedel
Quoted 7 lines
cpu_has() returns int,
but would it be better to have something like
if (!mce_disabled &&
	    !(c->x86_capability & (X86_FEATURE_MCA | X86_FEATURE_MCE)) {
		printk(KERN_INFO "CPU%i: No machine check support available\n",
			smp_processor_id());

Quoted 5 lines

Quoted 2 lines
This looks complicated and is harder to read. Its exactly the purpose of the cpu_has() macro to avoid such constructs.
It is done via test_bit(), which is designed for IO access with all that `const volatile' stuff, 2 x unnecessary, can't be optimized here (IMHO).

This code runs only on bootup and once per cpu. I think readable code is more important here than make it a few cycles faster.

Signature · 6 lines

           |           AMD Saxony Limited Liability Company & Co. KG
 Operating |         Wilschdorfer Landstr. 101, 01109 Dresden, Germany
 System    |                  Register Court Dresden: HRA 4896
 Research  |              General Partner authorized to represent:
 Center    |             AMD Saxony LLC (Wilmington, Delaware, US)
           | General Manager of AMD Saxony LLC: Dr. Hans-R. Deppe, Thomas McCoy

[email protected]

2007-10-09 20:47 UTC #6 lore raw

Re: [PATCH 1/2] i386: mce cleanup part1: functional change

In reply to Oleg Verych

Quoted 9 lines · Oleg Verych

Quoted 8 lines · Joerg Roedel
Quoted 7 lines
cpu_has() returns int,
but would it be better to have something like
if (!mce_disabled &&
	    !(c->x86_capability & (X86_FEATURE_MCA | X86_FEATURE_MCE)) {
		printk(KERN_INFO "CPU%i: No machine check support available\n",
			smp_processor_id());

Quoted 5 lines

Quoted 2 lines
This looks complicated and is harder to read. Its exactly the purpose of the cpu_has() macro to avoid such constructs.
It is done via test_bit(), which is designed for IO access with all that `const volatile' stuff, 2 x unnecessary, can't be optimized here (IMHO).

If this code is getting called often enough that optimization matters, you got *bigger* issues to worry about than optimization. Looks like it should only happen once at boot time.

Oleg Verych

2007-10-10 01:44 UTC #7 lore raw

Re: [PATCH 1/2] i386: mce cleanup part1: functional change

Quoted 13 lines · On Tue, Oct 09, 2007 at 04:46:46PM -0400, [email protected] wrote:

Quoted 12 lines · Oleg Verych
Quoted 10 lines · Joerg Roedel
Quoted 7 lines
cpu_has() returns int,
but would it be better to have something like
if (!mce_disabled &&
	    !(c->x86_capability & (X86_FEATURE_MCA | X86_FEATURE_MCE)) {
		printk(KERN_INFO "CPU%i: No machine check support available\n",
			smp_processor_id());
This looks complicated and is harder to read. Its exactly the purpose of the cpu_has() macro to avoid such constructs.

Quoted 6 lines

Quoted 2 lines
It is done via test_bit(), which is designed for IO access with all that `const volatile' stuff, 2 x unnecessary, can't be optimized here (IMHO).
If this code is getting called often enough that optimization matters, you got *bigger* issues to worry about than optimization. Looks like it should only happen once at boot time.

Text size matters even on static storages. A Linuxbios image not fitting to the 2M flash, etc.

Thanks.
____

Oleg Verych

2007-10-09 17:19 UTC #8 lore raw

coding for optimizations (Re: [PATCH 1/2] i386: mce cleanup part1: functional change)

In reply to Joerg Roedel

Quoted 25 lines · Joerg Roedel

Quoted 20 lines
Quoted 6 lines
void mcheck_init(struct cpuinfo_x86 *c)
{
arch/i386/kernel/cpu/mcheck/mce.c +2 -0
+	uint32_t mca, mce;
+
 	if (mce_disabled==1)
 		return;
arch/i386/kernel/cpu/mcheck/mce.c void fastcall (*machine_check_vector)(struct pt_regs *, long error_code) = unexp · line 49
Quoted 10 lines
arch/i386/kernel/cpu/mcheck/mce.c +9 -0
 
+	mca = cpu_has(c, X86_FEATURE_MCA);
+	mce = cpu_has(c, X86_FEATURE_MCE);
+
+	if (!mca || !mce) {
+		printk(KERN_INFO "CPU%i: No machine check support available\n",
+			smp_processor_id());
+		return;
+	}
+
cpu_has() returns int,
but would it be better to have something like
Quoted 4 lines
if (!mce_disabled &&
	    !(c->x86_capability & (X86_FEATURE_MCA | X86_FEATURE_MCE)) {
		printk(KERN_INFO "CPU%i: No machine check support available\n",
			smp_processor_id());

Quoted 8 lines

This looks complicated and is harder to read. Its exactly the purpose of the cpu_has() macro to avoid such constructs.
Quoted 3 lines
		return;
	} else
		return;
Return unconditionaly here?

Kind of typo.

Due to arch code, i think, it must be coded in non-gcc optimistic way. As practice and Linus shows, gcc's optimizations sometimes produce very unexpected results. People do spaghetti-like coding, without thinking about text size / run time.

Compile time payment was noted by Andrew many years ago:

"As you know, I'd be more concerned about moves to drop support for the older and much faster gcc versions. If you're not using egcs-1.1.2, you're already a very patient person." <http://lkml.org/lkml/2001/12/29/163>

So, i actually wanted to write this:

	if (!mce_disabled) {
	if (!(c->x86_capability & (X86_FEATURE_MCA | X86_FEATURE_MCE)) {
		printk(KERN_INFO "CPU%i: No machine check support available\n",
 			smp_processor_id());
		return;
	}
	/* function code */
	}

(ugly, because `mce_disabled == 1' check is even in gcc is likely by default). But changed my mind, due to violation of the coding style. Of course this my be no appropriate at all, but i'd like to bring this, if anyone would be like to discuss this.
____

Joerg Roedel

author

2007-10-09 18:31 UTC #9 lore raw

Re: coding for optimizations (Re: [PATCH 1/2] i386: mce cleanup part1: functional change)

Quoted 8 lines · On Tue, Oct 09, 2007 at 07:33:17PM +0200, [email protected] wrote:

	if (!mce_disabled) {
	if (!(c->x86_capability & (X86_FEATURE_MCA | X86_FEATURE_MCE)) {
		printk(KERN_INFO "CPU%i: No machine check support available\n",
 			smp_processor_id());
		return;
	}
	/* function code */
	}

I have 2 problems with this way:

1) It is totally broken because c->x86_capability is not an integer but an *array* of integers. This is also the reason why test_bit() is used in cpu_has() and not plain logical operators.

2) It is still hard to read and breaks the kernel coding style.

But you are right with the redundant mca and mce variables. They are not needed and I will inline the cpu_has() checks into the condition check. I'll resubmit tomorrow.

Signature · 6 lines

           |           AMD Saxony Limited Liability Company & Co. KG
 Operating |         Wilschdorfer Landstr. 101, 01109 Dresden, Germany
 System    |                  Register Court Dresden: HRA 4896
 Research  |              General Partner authorized to represent:
 Center    |             AMD Saxony LLC (Wilmington, Delaware, US)
           | General Manager of AMD Saxony LLC: Dr. Hans-R. Deppe, Thomas McCoy

Adrian Bunk

2007-10-10 23:14 UTC #10 lore raw

Re: coding for optimizations (Re: [PATCH 1/2] i386: mce cleanup part1: functional change)

Quoted 4 lines · Joerg Roedel

...
But you are right with the redundant mca and mce variables. They are not needed and I will inline the cpu_has() checks into the condition check. I'll resubmit tomorrow.

Please don't let Oley bring you away from the right path.

It is your job to write readable C code, and it's the compiler's job to transform this C code into efficient machine code.

And non-ancient gcc versions are usually quite good in optimizing code.

And saving code in a variable before using it might even result in the same assembler code.

It is personal preference whether you use variables or not in this case (both seem to be equally readable) so you can do it any way you like, but don't try to guess what gcc might make out of it.

We have problems with the size of the kernel image, but we need to solve them with the bigger knobs that have measurable effects, not by wasting our time on guessing how gcc might handle a single if.

It's also a quite ill idea to think about whether gcc might produce a few bytes more or less code at the if when there's such a long printk() in the middle...

cu
Adrian

Signature · 4 lines

       "Is there not promise of rain?" Ling Tan asked suddenly out
        of the darkness. There had been need of rain for many days.
       "Only a promise," Lao Er said.
                                       Pearl S. Buck - Dragon Seed

Oleg Verych

2007-10-11 15:12 UTC #11 lore raw

Re: coding for optimizations (Re: [PATCH 1/2] i386: mce cleanup part1: functional change)

On Thu, Oct 11, 2007 at 01:14:29AM +0200, Adrian Bunk wrote:
[]

Quoted 3 lines

It's also a quite ill idea to think about whether gcc might produce a few bytes more or less code at the if when there's such a long printk() in the middle...

printk() problem was discussed with proper banana userspace replacement proposition by me, so i don't care much.

Though, why MCE can't be enabled/disabled by config option? Native engineers from proper company should know what processors have this feature... If kconfig isn't flexible for this, i'm glad to hear opinions. If it's needed anyway, then sorry.
____

Adrian Bunk

2007-10-11 15:21 UTC #12 lore raw

Re: coding for optimizations (Re: [PATCH 1/2] i386: mce cleanup part1: functional change)

Quoted 6 lines · Oleg Verych

On Thu, Oct 11, 2007 at 01:14:29AM +0200, Adrian Bunk wrote:
[]
Quoted 3 lines
It's also a quite ill idea to think about whether gcc might produce a few bytes more or less code at the if when there's such a long printk() in the middle...

Quoted 7 lines

printk() problem was discussed with proper banana userspace replacement proposition by me, so i don't care much.
Though, why MCE can't be enabled/disabled by config option? Native engineers from proper company should know what processors have this feature... If kconfig isn't flexible for this, i'm glad to hear opinions. If it's needed anyway, then sorry.

You should better say sorry for not having checked that it can already be disabled with a kconfig option...

cu
Adrian

Signature · 4 lines

       "Is there not promise of rain?" Ling Tan asked suddenly out
        of the darkness. There had been need of rain for many days.
       "Only a promise," Lao Er said.
                                       Pearl S. Buck - Dragon Seed

Oleg Verych

2007-10-11 15:59 UTC #13 lore raw

Re: coding for optimizations (Re: [PATCH 1/2] i386: mce cleanup part1: functional change)

Quoted 10 lines · Adrian Bunk

Quoted 9 lines · Oleg Verych
On Thu, Oct 11, 2007 at 01:14:29AM +0200, Adrian Bunk wrote:
[]
Quoted 3 lines
It's also a quite ill idea to think about whether gcc might produce a few bytes more or less code at the if when there's such a long printk() in the middle...
printk() problem was discussed with proper banana userspace replacement proposition by me, so i don't care much.

Quoted 7 lines

Quoted 4 lines
Though, why MCE can't be enabled/disabled by config option? Native engineers from proper company should know what processors have this feature... If kconfig isn't flexible for this, i'm glad to hear opinions. If it's needed anyway, then sorry.
You should better say sorry for not having checked that it can already be disabled with a kconfig option...

Talk is not about wana-build-MCE-in, given to the user's sake. But about flexible selection of this code, if it is really supported by CPU, when user have it enabled (see winchip/preventium sub-thread). More work, it is not such quick patch, certainly, but this is needed for famous i386.

I wonder, why such patch wasn't introduced long time ago. Modern (cheap) chips have no MCE?
____

Joerg Roedel

author

2007-10-09 12:50 UTC #14 lore raw

PATCH 2/2 i386: mce cleanup part2: conding style cleanups

In reply to Joerg Roedel

Remove one indent level

Signed-off-by Christoph Egger <[email protected]>

Signed-off-by Joerg Roedel <[email protected]>

---
 arch/i386/kernel/cpu/mcheck/mce.c |   40 ++++++++++++++++++------------------
 1 files changed, 20 insertions(+), 20 deletions(-)

arch/i386/kernel/cpu/mcheck/mce.c +20 ~~-20~~

index bad2c64..fd743c3 100644
@@ -48,26 +48,26 @@ void mcheck_init(struct cpuinfo_x86 *c)
 	}
 
 	switch (c->x86_vendor) {
-		case X86_VENDOR_AMD:
-			amd_mcheck_init(c);
-			break;
-
-		case X86_VENDOR_INTEL:
-			if (c->x86==5)
-				intel_p5_mcheck_init(c);
-			if (c->x86==6)
-				intel_p6_mcheck_init(c);
-			if (c->x86==15)
-				intel_p4_mcheck_init(c);
-			break;
-
-		case X86_VENDOR_CENTAUR:
-			if (c->x86==5)
-				winchip_mcheck_init(c);
-			break;
-
-		default:
-			break;
+	case X86_VENDOR_AMD:
+		amd_mcheck_init(c);
+		break;
+
+	case X86_VENDOR_INTEL:
+		if (c->x86==5)
+			intel_p5_mcheck_init(c);
+		if (c->x86==6)
+			intel_p6_mcheck_init(c);
+		if (c->x86==15)
+			intel_p4_mcheck_init(c);
+		break;
+
+	case X86_VENDOR_CENTAUR:
+		if (c->x86==5)
+			winchip_mcheck_init(c);
+		break;
+
+	default:
+		break;
 	}
 }

Signature · 1 line

1.5.2.5