On Tue, 2014-09-30 at 12:09 +0200, Borislav Petkov wrote:
> On Tue, Sep 30, 2014 at 05:56:31PM +0800, Chen Yucong wrote:
> > I just clear it to avoid that the mce_log() call logs the above
> > threshold event again in machine_check_poll().
>
> Ok, that's a good point, please put it in the commit message.
>
> > It is just used for scanning other banks for recording other valid
> > error information.
>
> This is actually not what we want - we want to log the errors which
> cause the overflow first and then the rest. So you don't need the goto
> but simply have the machine_check_poll() at the end.
From: Chen Yucong <[email protected]>
machine_check_poll() will reset IA32_MCi_STATUS register to zero.
So we need to save the content of IA32_MCi_STATUS MSRs before
calling machine_check_poll() for logging threshold interrupt event.
mce_setup() does not gather the content of IA32_MCG_STATUS, so it
should be read explicitly. Moreover, we need to clear IA32_MCx_STATUS
to avoid that mce_log() logs the processed threshold event again
at next time.
Signed-off-by: Chen Yucong <[email protected]>
---
arch/x86/kernel/cpu/mcheck/mce_amd.c | 18 +++++++++++-------
1 file changed, 11 insertions(+), 7 deletions(-)
diff --git a/arch/x86/kernel/cpu/mcheck/mce_amd.c b/arch/x86/kernel/cpu/mcheck/mce_amd.c
index f8c56bd..643e6a2 100644
--- a/arch/x86/kernel/cpu/mcheck/mce_amd.c
+++ b/arch/x86/kernel/cpu/mcheck/mce_amd.c
@@ -274,6 +274,7 @@ static void amd_threshold_interrupt(void)
struct mce m;
mce_setup(&m);
+ rdmsrl(MSR_IA32_MCG_STATUS, m.mcgstatus);
/* assume first bank caused it */
for (bank = 0; bank < mca_cfg.banks; ++bank) {
@@ -305,24 +306,27 @@ static void amd_threshold_interrupt(void)
(high & MASK_LOCKED_HI))
continue;
- /*
- * Log the machine check that caused the threshold
- * event.
- */
- machine_check_poll(MCP_TIMESTAMP,
- this_cpu_ptr(&mce_poll_banks));
-
if (high & MASK_OVERFLOW_HI) {
rdmsrl(address, m.misc);
rdmsrl(MSR_IA32_MCx_STATUS(bank), m.status);
+ if (m.status & MCI_STATUS_ADDRV)
+ rdmsrl(MSR_IA32_MCx_ADDR(bank), m.addr);
m.bank = K8_MCE_THRESHOLD_BASE
+ bank * NR_BLOCKS
+ block;
mce_log(&m);
+
+ wrmsrl(MSR_IA32_MCx_STATUS(bank), 0);
return;
}
}
}
+
+ /*
+ * Log the machine check that caused the threshold event.
+ */
+ machine_check_poll(MCP_TIMESTAMP,
+ this_cpu_ptr(&mce_poll_banks));
}
/*
--
1.7.10.4
On Wed, Oct 01, 2014 at 12:35:02PM +0800, Chen Yucong wrote:
> On Tue, 2014-09-30 at 12:09 +0200, Borislav Petkov wrote:
> > On Tue, Sep 30, 2014 at 05:56:31PM +0800, Chen Yucong wrote:
> > > I just clear it to avoid that the mce_log() call logs the above
> > > threshold event again in machine_check_poll().
> >
> > Ok, that's a good point, please put it in the commit message.
> >
> > > It is just used for scanning other banks for recording other valid
> > > error information.
> >
> > This is actually not what we want - we want to log the errors which
> > cause the overflow first and then the rest. So you don't need the goto
> > but simply have the machine_check_poll() at the end.
>
>
> From: Chen Yucong <[email protected]>
>
> machine_check_poll() will reset IA32_MCi_STATUS register to zero.
> So we need to save the content of IA32_MCi_STATUS MSRs before
> calling machine_check_poll() for logging threshold interrupt event.
>
> mce_setup() does not gather the content of IA32_MCG_STATUS, so it
> should be read explicitly. Moreover, we need to clear IA32_MCx_STATUS
> to avoid that mce_log() logs the processed threshold event again
> at next time.
>
> Signed-off-by: Chen Yucong <[email protected]>
> ---
> arch/x86/kernel/cpu/mcheck/mce_amd.c | 18 +++++++++++-------
> 1 file changed, 11 insertions(+), 7 deletions(-)
>
> diff --git a/arch/x86/kernel/cpu/mcheck/mce_amd.c b/arch/x86/kernel/cpu/mcheck/mce_amd.c
> index f8c56bd..643e6a2 100644
> --- a/arch/x86/kernel/cpu/mcheck/mce_amd.c
> +++ b/arch/x86/kernel/cpu/mcheck/mce_amd.c
> @@ -274,6 +274,7 @@ static void amd_threshold_interrupt(void)
> struct mce m;
>
> mce_setup(&m);
> + rdmsrl(MSR_IA32_MCG_STATUS, m.mcgstatus);
>
> /* assume first bank caused it */
> for (bank = 0; bank < mca_cfg.banks; ++bank) {
> @@ -305,24 +306,27 @@ static void amd_threshold_interrupt(void)
> (high & MASK_LOCKED_HI))
> continue;
>
> - /*
> - * Log the machine check that caused the threshold
> - * event.
> - */
> - machine_check_poll(MCP_TIMESTAMP,
> - this_cpu_ptr(&mce_poll_banks));
> -
> if (high & MASK_OVERFLOW_HI) {
> rdmsrl(address, m.misc);
> rdmsrl(MSR_IA32_MCx_STATUS(bank), m.status);
> + if (m.status & MCI_STATUS_ADDRV)
> + rdmsrl(MSR_IA32_MCx_ADDR(bank), m.addr);
> m.bank = K8_MCE_THRESHOLD_BASE
> + bank * NR_BLOCKS
> + block;
> mce_log(&m);
> +
> + wrmsrl(MSR_IA32_MCx_STATUS(bank), 0);
> return;
Ok, this return is still bugging me - we're logging the error which
caused the counter overflow but we go and explicitly clear _STATUS so
that machine_check_poll doesn't pick up the same error again.
Even though, machine_check_poll is intended to log the thresholding
error.
Which actually makes me think that that machine_check_poll is actually
completely useless there. IOW, how about that instead:
---
From: Chen Yucong <[email protected]>
Date: Thu, 2 Oct 2014 14:48:19 +0200
Subject: [PATCH] x86, MCE, AMD: Correct thresholding error logging
mce_setup() does not gather the content of IA32_MCG_STATUS, so it
should be read explicitly. Moreover, we need to clear IA32_MCx_STATUS
to avoid that mce_log() logs the processed threshold event again
at next time.
But we do the logging ourselves and machine_check_poll() is completely
useless there. So kill it.
Signed-off-by: Chen Yucong <[email protected]>
Signed-off-by: Borislav Petkov <[email protected]>
---
arch/x86/kernel/cpu/mcheck/mce_amd.c | 30 +++++++++++++++---------------
1 file changed, 15 insertions(+), 15 deletions(-)
diff --git a/arch/x86/kernel/cpu/mcheck/mce_amd.c b/arch/x86/kernel/cpu/mcheck/mce_amd.c
index 1c54d3d61a4d..9ce64955559d 100644
--- a/arch/x86/kernel/cpu/mcheck/mce_amd.c
+++ b/arch/x86/kernel/cpu/mcheck/mce_amd.c
@@ -270,14 +270,13 @@ void mce_amd_feature_init(struct cpuinfo_x86 *c)
static void amd_threshold_interrupt(void)
{
u32 low = 0, high = 0, address = 0;
+ int cpu = smp_processor_id();
unsigned int bank, block;
struct mce m;
- mce_setup(&m);
-
/* assume first bank caused it */
for (bank = 0; bank < mca_cfg.banks; ++bank) {
- if (!(per_cpu(bank_map, m.cpu) & (1 << bank)))
+ if (!(per_cpu(bank_map, cpu) & (1 << bank)))
continue;
for (block = 0; block < NR_BLOCKS; ++block) {
if (block == 0) {
@@ -309,20 +308,21 @@ static void amd_threshold_interrupt(void)
* Log the machine check that caused the threshold
* event.
*/
- machine_check_poll(MCP_TIMESTAMP,
- &__get_cpu_var(mce_poll_banks));
-
- if (high & MASK_OVERFLOW_HI) {
- rdmsrl(address, m.misc);
- rdmsrl(MSR_IA32_MCx_STATUS(bank), m.status);
- m.bank = K8_MCE_THRESHOLD_BASE
- + bank * NR_BLOCKS
- + block;
- mce_log(&m);
- return;
- }
+ if (high & MASK_OVERFLOW_HI)
+ goto log;
}
}
+ return;
+
+log:
+ mce_setup(&m);
+ rdmsrl(MSR_IA32_MCG_STATUS, m.mcgstatus);
+ rdmsrl(address, m.misc);
+ rdmsrl(MSR_IA32_MCx_STATUS(bank), m.status);
+ m.bank = K8_MCE_THRESHOLD_BASE + bank * NR_BLOCKS + block;
+ mce_log(&m);
+
+ wrmsrl(MSR_IA32_MCx_STATUS(bank), 0);
}
/*
--
2.0.0
--
Regards/Gruss,
Boris.
Sent from a fat crate under my desk. Formatting is fine.
--
On Thu, 2014-10-02 at 15:12 +0200, Borislav Petkov wrote:
>
> Ok, this return is still bugging me - we're logging the error which
> caused the counter overflow but we go and explicitly clear _STATUS so
> that machine_check_poll doesn't pick up the same error again.
>
> Even though, machine_check_poll is intended to log the thresholding
> error.
>
> Which actually makes me think that that machine_check_poll is actually
> completely useless there. IOW, how about that instead:
amd_threshold_interrup() is just used for logging threshold events. And
any 'valid' threshold events can be checked/logged in loop body.
Moreover, machine_check_poll() is unable to check additional MCx_MISCi.
So I agree with you on this change.
Thanks!
cyc