On Thu, 21 Sep 2017 19:57:20 +1000
Michael Neuling [off-list ref] wrote:

On Thu, 2017-09-21 at 18:18 +1000, Nicholas Piggin wrote:

quoted

On Thu, 21 Sep 2017 12:04:34 +1000
Michael Neuling [off-list ref] wrote:
  

quoted

On POWER9 DD2.1 and below, it's possible to get Machine Check
Exception (MCE) where only DSISR bit 33 is set. This will result in
the linux MCE handler seeing an unknown event, which triggers linux to
crash.

We change this by detecting unknown events in the MCE handler and
marking them as handled so that we no longer crash. We do this only on
chip revisions known to have this problem.

MCE that occurs like this is spurious, so we don't need to do anything
in terms of servicing it. If there is something that needs to be
serviced, the CPU will raise the MCE again with the correct DSISR so
that it can be serviced properly.

Signed-off-by: Michael Neuling <redacted>
---
v2 update commit message based on Balbir's comments
---
 arch/powerpc/kernel/mce_power.c | 15 +++++++++++++++
 1 file changed, 15 insertions(+)

diff --git a/arch/powerpc/kernel/mce_power.c

b/arch/powerpc/kernel/mce_power.c
index b76ca198e0..72ec667136 100644

--- a/arch/powerpc/kernel/mce_power.c
+++ b/arch/powerpc/kernel/mce_power.c

@@ -595,6 +595,7 @@ static long mce_handle_error(struct pt_regs *regs,

 	uint64_t addr;
 	uint64_t srr1 = regs->msr;
 	long handled;
+	unsigned long pvr;
 
 	if (SRR1_MC_LOADSTORE(srr1))
 		handled = mce_handle_derror(regs, dtable, &mce_err, &addr);

@@ -604,6 +605,20 @@ static long mce_handle_error(struct pt_regs *regs,

 	if (!handled && mce_err.error_type == MCE_ERROR_TYPE_UE)
 		handled = mce_handle_ue_error(regs);
 
+	/*
+	 * On POWER9 DD2.1 and below, it's possible to get machine
+	 * check where only DSISR bit 33 is set. This will result in
+	 * the MCE handler seeing an unknown event and us crashing.
+	 * Change this to mark as handled on these revisions.
+	 */
+	pvr = mfspr(SPRN_PVR);
+	if (((PVR_VER(pvr) == PVR_POWER9) &&
+	     (PVR_CFG(pvr) == 2) &&
+	     (PVR_MIN(pvr) <= 1)) || cpu_has_feature(CPU_FTR_POWER9_DD1))
+		/* DD2.1 and below */
+		if (mce_err.error_type == MCE_ERROR_TYPE_UNKNOWN)
+		    handled = 1;  

I might be missing something, but can you just do

  if (regs->dsisr == 0x40000000)
      return 1;

In __machine_check_early_realmode_p9() ?  

You're right, thanks.

If you leave the PVR and DD1 checks in there, it would be a good
reminder for me to convert into a quirk if I can get this version
specific quirks stuff going

https://marc.info/?l=linuxppc-embedded&m=150597337720114&w=2

Thanks,
Nick