On Jan 9, 2016 11:51 AM, "Tony Luck" <tony.luck(a)gmail.com> wrote:
> Oh, I see. Is it the case that the MC code can't cleanly handle the
> case where the error was nominally recoverable but the kernel doesn't
> know how to recover from it due to the lack of a handler that's okay
> with it, because the handler's refusal to handle the fault wouldn't be
> known until too late?
The code is just too clunky right now. We have a table driven
severity calculator that we invoke on each machine check bank
that has some valid data to report. Part of that calculation is
"what context am I in?". Which happens earlier in the sequence
than "Is MCi_STATUS.MCACOD some known recoverable type".
If I invoke the fixup code I'll change regs->ip right away ... even
if I'm executing on some innocent bystander processor that wasn't
the source of the machine check (the bystanders on the same
socket can usually see something logged in one of the memory
Makes sense, sort of. But even if there is an MC fixup registered,
don't you still have to make sure to execute it on the actual victim
CPU? After all, you don't want to fail an mcsafe copy just because a
different CPU coincidentally machine checked while the mcsafe copy has
the recoverable RIP value.
There are definitely some cleanups that should be done
in this code (e.g. figuring our context just once, not once
per bank). But I'm pretty sure I'll always want to know
"am I executing an instruction with a #MC recoverable
handler?" in a way that doesn't actually invoke the recovery.
What's wrong with:
Step 1: determine that the HW context is, in principle, recoverable.
Step 2: ask the handler to try to recover.
Step 3: if the handler doesn't recover, panic
I'm not saying that restructuring the code like this should be a
prerequisite for merging this, but I'm wondering whether it would make
sense at some point in the future.