Unrecognized MSR report
A message that is not in the archive (sent off-list or not yet received): [email protected]
Borislav Petkov
maintainer+ Mauro, Tony, linux-edac
Quoted 9 lines · Andres Valloud
Hello, after installing rasdaemon the logs complain thus.
[604477.995036] msr: Write to unrecognized MSR 0x17f by rasdaemon (pid:24480).
The hardware in question is a supermicro server with 2x E5-2630v2. A newer server with 2x Xeon Gold 6128 is also running rasdaemon, but there is no complaint about MSR
0x17fthere.
Quoted 7 lines
Both systems are running Ubuntu 22.04 LTS, fully up to date, with kernel version 6.5.0-44.
The text in the MSR document referenced from the logs only makes reference to AMD cpus. There is also the different behavior depending on the Intel cpu model. This circumstance does not appear to be covered by the document. Reporting as requested...
Thanks, appreciated.
---
From: "Borislav Petkov (AMD)" <[email protected]>
Date: Fri, 26 Jul 2024 14:39:53 +0200
Subject: [PATCH] mce-intel: Do not enable optional MCE logging in rasdaemon
The kernel already does that, see kernel commit:
68299a42f842 ("x86/mce: Enable additional error logging on certain Intel CPUs")
so rip out yet another case of poking into MSRs from userspace.
--- mce-intel.c | 26 -------------------------- ras-mce-handler.c | 11 ----------- ras-mce-handler.h | 3 --- 3 files changed, 40 deletions(-)
mce-intel.c +0 -26
@@ -461,29 +461,3 @@ static int domsr(int cpu, int msr, int bit)close (fd);return 0 ; } - -int set_intel_imc_log (enum cputype cputype,unsigned int ncpus) -{ -int cpu, msr, bit, rc; - -switch (cputype) { -case CPU_SANDY_BRIDGE_EP : -case CPU_IVY_BRIDGE_EPEX : -case CPU_HASWELL_EPEX : -case CPU_KNIGHTS_LANDING : -case CPU_KNIGHTS_MILL : - msr =0x17f ;/* MSR_ERROR_CONTROL */ - bit =0x2 ;/* MemError Log Enable */ -break ; -default : -return 0 ; - } - -for (cpu =0 ; cpu < ncpus; cpu++) { - rc =domsr (cpu, msr, bit); -if (rc) -return rc; - } - -return 0 ; -}
ras-mce-handler.c +0 -11
@@ -249,17 +249,6 @@ int register_mce_handler(struct ras_events *ras, unsigned int ncpus) ras->mce_priv =NULL ;return rc; } -switch (mce->cputype) { -case CPU_SANDY_BRIDGE_EP : -case CPU_IVY_BRIDGE_EPEX : -case CPU_HASWELL_EPEX : -case CPU_KNIGHTS_LANDING : -case CPU_KNIGHTS_MILL : -set_intel_imc_log (mce->cputype, ncpus); -default : -break ; - } -return rc; }
ras-mce-handler.h +0 -3
@@ -109,9 +109,6 @@ int ras_mce_event_handler(struct trace_seq *s,struct tep_record *record, struct tep_event *event, void *context); -/* enables intel iMC logs */ -int set_intel_imc_log (enum cputype cputype,unsigned int ncpus); -/* Undertake AMD SMCA Error Decoding */ void decode_smca_error (struct mce_event *e,struct mce_priv *m);void amd_decode_errcode (struct mce_event *e);
Signature · 7 lines
2.43.0
--
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquetteQuoted 7 lines
From: Borislav Petkov <[email protected]>
Sent: Friday, July 26, 2024 8:43 PM
To: Andres Valloud <[email protected]>; Mauro Carvalho Chehab <[email protected]>; Luck, Tony <[email protected]>
Cc: [email protected]; linux-edac <[email protected]>
Subject: Re: Unrecognized MSR report
Quoted 8 lines
+ Mauro, Tony, linux-edac
Quoted 5 lines · Andres Valloud
Hello, after installing rasdaemon the logs complain thus.
[604477.995036] msr: Write to unrecognized MSR 0x17f by rasdaemon (pid:24480).
The hardware in question is a supermicro server with 2x E5-2630v2. A
E5-2630 is the Intel CPU with the codename "Ivy Bridge EP" with MSR 0x17F capability.
I reproduced the kernel warning message above on an Intel SandyBridge, This was because of the default sys-node:
/sys/module/msr/parameters/allow_writes == "default"
If I set this sys-node to
/sys/module/msr/parameters/allow_writes = "on"
Then there was no such warning kernel message.
Quoted 2 lines
Quoted 2 lines
newer server with 2x Xeon Gold 6128 is also running rasdaemon, but there is no complaint about MSR
0x17fthere.
Xeon Gold 6128 is the Intel CPU with the code name "Skylake".
As Skylake didn't have MSR 0x17f and rasdaemon didn't try to write this MSR, so no kernel warning message on this CPU.
Quoted 16 lines
Quoted 6 lines
Both systems are running Ubuntu 22.04 LTS, fully up to date, with kernel version 6.5.0-44.
The text in the MSR document referenced from the logs only makes reference to AMD cpus. There is also the different behavior depending on the Intel cpu model. This circumstance does not appear to be covered
by the document.
Quoted 1 line
Reporting as requested...
Thanks, appreciated.
---
From: "Borislav Petkov (AMD)" <[email protected]>
Date: Fri, 26 Jul 2024 14:39:53 +0200
Subject: [PATCH] mce-intel: Do not enable optional MCE logging in rasdaemon
Quoted 6 lines
The kernel already does that, see kernel commit:
68299a42f842("x86/mce: Enable additional error logging on certain Intel
CPUs")
so rip out yet another case of poking into MSRs from userspace.
Just a small concern:
The commit 68299a42f842 was merged to kernel v5.11.
For kernels < v5.11 and w/o backporting this commit may still need the rasdaemon to wrmsr 0x17f to enable the additional error logging for these certain Intel CPUs.
If get the rasdameon to write the sys-node as below before doing wrmsr 0x17f for these CPUs
/sys/module/msr/parameters/allow_writes = "on"
Then there will be no such kernel warning message:
msr: Write to unrecognized MSR 0x17f by rasdaemon (pid: 24480).
@Borislav Petkov, if you think the fix above to the rasdaemon is acceptable, I'll proceed to make a fix for rasdaemon.
Quoted 4 lines
Reported-by Andres Valloud <[email protected]>Signed-off-by Borislav Petkov (AMD) <[email protected]>---
[...]
Thanks!
-Qiuxu
Borislav Petkov
maintainerQuoted 10 lines · Qiuxu
I reproduced the kernel warning message above on an Intel SandyBridge, This was because of the default sys-node:
/sys/module/msr/parameters/allow_writes == "default"If I set this sys-node to
/sys/module/msr/parameters/allow_writes= "on"Then there was no such warning kernel message.
Because poking at MSRs from userspace is a bad idea.
Quoted 5 lines
Just a small concern:
The commit
68299a42f842was merged to kernel v5.11.
For kernels < v5.11 and w/o backporting this commit may still need the rasdaemon to wrmsr0x17fto enable the additional error logging for these certain Intel CPUs.
What user runs an old kernel (< 5.11) with a new rasdaemon?
If there's a valid use case and they've backported stuff, then they can backport the above commit too.
Quoted 10 lines
If get the rasdameon to write the sys-node as below before doing wrmsr
0x17ffor these CPUs
/sys/module/msr/parameters/allow_writes= "on"Then there will be no such kernel warning message:
msr: Write to unrecognized MSR 0x17f by rasdaemon (pid: 24480).@Borislav Petkov, if you think the fix above to the rasdaemon is acceptable, I'll proceed to make a fix for rasdaemon.
The goal is to stop poking at MSRs from userspace. And I've sent a patch already.
What fix are you going to proceed with?
Signature · 4 lines
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquetteQuoted 3 lines
From: Borislav Petkov <[email protected]>
[...]
Because poking at MSRs from userspace is a bad idea.
Yes.
Quoted 2 lines
[...]
What user runs an old kernel (< 5.11) with a new rasdaemon?
I thought some cloud service providers running old kernels (<5.11) could upgrade their rasdaemons at runtime (w/o system reboot).
Quoted 2 lines
If there's a valid use case and they've backported stuff, then they can backport the above commit too.
Yes, agree.
Quoted 8 lines
[...]
Quoted 2 lines
@Borislav Petkov, if you think the fix above to the rasdaemon is acceptable, I'll proceed to make a fix for rasdaemon.
The goal is to stop poking at MSRs from userspace. And I've sent a patch already.
What fix are you going to proceed with?
Some pseudo-codes like below.
However, as you mentioned, poking MSR in userspace is not advisable. And I agree that users can backport that kernel commit if they use a new rasdaemon. Please go ahead with yours 😊.
--------
mce-intel.c +7 -1
@@ -486,11 +486,17 @@ int set_intel_imc_log(enum cputype cputype, unsigned int ncpus)return 0 ; } + old_state =get_wrmsr_state ("/sys/module/msr/parameters/allow_writes" ); +set_wrmsr_state ("/sys/module/msr/parameters/allow_writes" ,"on" ); +for (cpu =0 ; cpu < ncpus; cpu++) { rc =domsr (cpu, msr, bit); -if (rc) +if (rc) { +set_wrmsr_state ("/sys/module/msr/parameters/allow_writes" , old_state);return rc; + } } +set_wrmsr_state ("/sys/module/msr/parameters/allow_writes" , old_state);return 0 ; }
Borislav Petkov
maintainerQuoted 2 lines · Qiuxu
I thought some cloud service providers running old kernels (<5.11) could upgrade their rasdaemons at runtime (w/o system reboot).
You thought. So no real use case. No, we don't anticipate use cases.
Quoted 2 lines
Some pseudo-codes like below. However, as you mentioned, poking MSR in userspace is not advisable.
You got it: we don't want to perpetuate the MSR poking - we want to get rid of it.
Thx.
Signature · 4 lines
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquette