Re: [PATCH v4 2/5] x86/msr: Carry on after a non-"safe" MSR access fails without !panic_on_oops
From: Andy Lutomirski
Date: Mon Mar 14 2016 - 14:24:27 EST
On Mon, Mar 14, 2016 at 11:15 AM, Linus Torvalds
> On Mon, Mar 14, 2016 at 11:10 AM, Andy Lutomirski <luto@xxxxxxxxxxxxxx> wrote:
>> A couple of the wrmsr users actually care about performance. These
>> are the ones involved in context switching and, to a lesser extent, in
>> switching in and out of guest mode.
> .. ok, see the crossed emails.
> I'd *much* rather special case the special cases. Not make the generic
> case something complex.
The code in my queue is, literally:
bool ex_handler_rdmsr_unsafe(const struct exception_table_entry *fixup,
struct pt_regs *regs, int trapnr)
WARN_ONCE(1, "unchecked MSR access error: RDMSR from 0x%x",
/* Pretend that the read succeeded and returned 0. */
regs->ip = ex_fixup_addr(fixup);
regs->ax = 0;
regs->dx = 0;
The only regard in which this is any more complex than the _safe
variant is because there's a warning (one line of code) and because
the _safe variant forgot to zero the result (two lines of code). My
series fixes the latter, so we're talking about almost no source code
size difference. There *is* a difference in binary size, though --
the _safe variant emits a copy of its fixup every time it appears,
whereas the new fixup appears once.
So I think we should apply my patches (with the early handling fixed
and the panic_on_oops removed), and then consider reimplementing the
_safe variant using fancy handlers to reduce number of lines of asm
and code size, and then consider changing the overall API on top if we
think there's a better API to be had.
Is that okay?
> And *particularly* not make the generic case be something where people
> think it's sane to oops and kill the machine. Christ.
I've already removed the panic_on_oops thing from my tree.