mce: include cmci during intel feature clearing
Subsystems (3)
PATCH 1/1 mce: include cmci during intel feature clearing
It was found that after a kexec on an intel CPU, MCE reporting was no longer active. The root cause has been found to be that ownership of CMCI banks is not cleared during the shutdown phase. As a result, when CPU's come back online, they are unable to rediscover these occupied banks. If we clear these CPU associations before booting into the new kernel, the CMCI banks can be reclaimed and MCE reporting will become functional once more.
The existing code does seem to have the intention of clearing MCE-related features via mcheck_cpu_clear(). During a kexec reboot, there are two sequences that reach a call to mcheck_cpu_clear(). They are:
1) Stopping other (remote) CPU's via IPI:native_machine_shutdown()
stop_other_cpus() smp_ops.stop_other_cpus(1) x86 smp: native_stop_other_cpus(1) apic_send_IPI_allbutself(REBOOT_VECTOR)
...IPI is received on remote CPU's and IDT sysvec_reboot invoked:
stop_this_cpu() mcheck_cpu_clear(this_ptr_cpu(&cpu_info))
2) Seqence of stopping the active CPU (the one performing the kexec):native_machine_shutdown()
stop_other_cpus() smp_ops.stop_other_cpus(1) x86 smp: native_stop_other_cpus(1) mcheck_cpu_clear(this_ptr_cpu(&cpu_info))
In both cases, the call to mcheck_cpu_clear() leads to the vendor specific call to intel_feature_clear():
mcheck_cpu_clear(this_ptr_cpu(&cpu_info))
__mcheck_cpu_clear_vendor(c) switch (c->x86_vendor) case X86_VENDOR_INTEL: mce_intel_feature_clear(c)
Now looking at the pair of functions mce_intel_feature_{init,clear}, there are 3 MCE features setup on the init side:
mce_intel_feature_init(c)
intel_init_cmci()
intel_init_lmce()
intel_imc_init(c)
On the other side in the clear function, only one of these features is actually cleared:
mce_intel_feature_clear(c)
intel_clear_lmce()
Just focusing on the feature pertaining to the root cause of the kexec issue, there would be a benefit if we additionally cleared the CMCI feature within this routine - the banks would be free for acquisition on the boot up side of a kexec. This patch adds the call to clear CMCI to this intel routine.
--- arch/x86/kernel/cpu/mce/intel.c | 1 + 1 file changed, 1 insertion(+)
arch/x86/kernel/cpu/mce/intel.c +1 -0
@@ -478,6 +478,7 @@ void mce_intel_feature_init(struct cpuinfo_x86 *c)void mce_intel_feature_clear (struct cpuinfo_x86 *c) {intel_clear_lmce (); +cmci_clear (); }bool intel_filter_mce (struct mce *m)
Signature · 1 line
2.47.1
Quoted 52 lines · JP Kobryn
It was found that after a kexec on an intel CPU, MCE reporting was no longer active. The root cause has been found to be that ownership of CMCI banks is not cleared during the shutdown phase. As a result, when CPU's come back online, they are unable to rediscover these occupied banks. If we clear these CPU associations before booting into the new kernel, the CMCI banks can be reclaimed and MCE reporting will become functional once more.
The existing code does seem to have the intention of clearing MCE-related features via
mcheck_cpu_clear(). During a kexec reboot, there are two sequences that reach a call tomcheck_cpu_clear(). They are:1) Stopping other (remote) CPU's via IPI:
native_machine_shutdown()stop_other_cpus() smp_ops.stop_other_cpus(1) x86 smp: native_stop_other_cpus(1) apic_send_IPI_allbutself(REBOOT_VECTOR)...IPI is received on remote CPU's and IDT
sysvec_rebootinvoked:stop_this_cpu() mcheck_cpu_clear(this_ptr_cpu(&cpu_info))2) Seqence of stopping the active CPU (the one performing the kexec):
native_machine_shutdown()stop_other_cpus() smp_ops.stop_other_cpus(1) x86 smp: native_stop_other_cpus(1) mcheck_cpu_clear(this_ptr_cpu(&cpu_info))In both cases, the call to
mcheck_cpu_clear()leads to the vendor specific call tointel_feature_clear():mcheck_cpu_clear(this_ptr_cpu(&cpu_info))
__mcheck_cpu_clear_vendor(c) switch (c->x86_vendor) case X86_VENDOR_INTEL: mce_intel_feature_clear(c)Now looking at the pair of functions mce_intel_feature_{init,clear}, there are 3 MCE features setup on the init side:
mce_intel_feature_init(c)
intel_init_cmci()
intel_init_lmce()
intel_imc_init(c)On the other side in the clear function, only one of these features is actually cleared:
mce_intel_feature_clear(c)
intel_clear_lmce()
Quoted 8 lines
Just focusing on the feature pertaining to the root cause of the kexec issue, there would be a benefit if we additionally cleared the CMCI feature within this routine - the banks would be free for acquisition on the boot up side of a kexec. This patch adds the call to clear CMCI to this intel routine.
Reported-by Aijay Adams <[email protected]>Signed-off-by JP Kobryn <[email protected]>
LGTM
arch/x86/kernel/cpu/mce/intel.c void mce_intel_feature_init(struct cpuinfo_x86 *c) · line 481
Quoted 13 lines
--- arch/x86/kernel/cpu/mce/intel.c | 1 + 1 file changed, 1 insertion(+)arch/x86/kernel/cpu/mce/intel.c +1
-0@@ -478,6 +478,7 @@ void mce_intel_feature_init(struct cpuinfo_x86 *c)void mce_intel_feature_clear (struct cpuinfo_x86 *c) {intel_clear_lmce (); +cmci_clear ();
I particularly like that you found a one-line fix!
Quoted 6 lines
arch/x86/kernel/cpu/mce/intel.c +0
-0}bool intel_filter_mce (struct mce *m)Signature · 1 line
2.47.1
Borislav Petkov
maintainerQuoted 49 lines · JP Kobryn
It was found that after a kexec on an intel CPU, MCE reporting was no longer active. The root cause has been found to be that ownership of CMCI banks is not cleared during the shutdown phase. As a result, when CPU's come back online, they are unable to rediscover these occupied banks. If we clear these CPU associations before booting into the new kernel, the CMCI banks can be reclaimed and MCE reporting will become functional once more.
The existing code does seem to have the intention of clearing MCE-related features via
mcheck_cpu_clear(). During a kexec reboot, there are two sequences that reach a call tomcheck_cpu_clear(). They are:1) Stopping other (remote) CPU's via IPI:
native_machine_shutdown()stop_other_cpus() smp_ops.stop_other_cpus(1) x86 smp: native_stop_other_cpus(1) apic_send_IPI_allbutself(REBOOT_VECTOR)...IPI is received on remote CPU's and IDT
sysvec_rebootinvoked:stop_this_cpu() mcheck_cpu_clear(this_ptr_cpu(&cpu_info))2) Seqence of stopping the active CPU (the one performing the kexec):
native_machine_shutdown()stop_other_cpus() smp_ops.stop_other_cpus(1) x86 smp: native_stop_other_cpus(1) mcheck_cpu_clear(this_ptr_cpu(&cpu_info))In both cases, the call to
mcheck_cpu_clear()leads to the vendor specific call tointel_feature_clear():mcheck_cpu_clear(this_ptr_cpu(&cpu_info))
__mcheck_cpu_clear_vendor(c) switch (c->x86_vendor) case X86_VENDOR_INTEL: mce_intel_feature_clear(c)Now looking at the pair of functions mce_intel_feature_{init,clear}, there are 3 MCE features setup on the init side:
mce_intel_feature_init(c)
intel_init_cmci()
intel_init_lmce()
intel_imc_init(c)On the other side in the clear function, only one of these features is actually cleared:
Quoted 8 lines
mce_intel_feature_clear(c)
intel_clear_lmce()Just focusing on the feature pertaining to the root cause of the kexec issue, there would be a benefit if we additionally cleared the CMCI feature within this routine - the banks would be free for acquisition on the boot up side of a kexec. This patch adds the call to clear CMCI to this intel routine.
Please:
- shorten this commit message - there really is no need to explain in such detail that mcheck_cpu_clear() has simply forgotten to clear CMCI banks too.
- run it through a spellchecker
- drop all personal pronouns
- write it in imperative tone
Some hints:
Section "2) Describe your changes" in Documentation/process/submitting-patches.rst for more details.
Also, see section "Changelog" inDocumentation/process/maintainer-tip.rst
Thx.
Signature · 4 lines
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquetteZhuo, Qiuxu
Reviewed-byHi JP,
Quoted 19 lines
[...]
Quoted 5 lines
Just focusing on the feature pertaining to the root cause of the kexec issue, there would be a benefit if we additionally cleared the CMCI feature within this routine - the banks would be free for acquisition on the boot up side of a kexec. This patch adds the call to clear CMCI to this intel routine.
Please:
- shorten this commit message - there really is no need to explain in such detail that
mcheck_cpu_clear()has simply forgotten to clear CMCI banks too.- run it through a spellchecker
- drop all personal pronouns
- write it in imperative tone
Quoted 8 lines
Some hints:
Section "2) Describe your changes" in
Documentation/process/submitting-patches.rstfor more details.Also, see section "Changelog" in
Documentation/process/maintainer-tip.rst
Please address Boris' comments, other than that