[PATCH] riscv: keep interrupts disabled for BREAKPOINT exception

[Masami Hiramatsu]

On Thu, 8 Apr 2021 19:23:48 +0800
Jisheng Zhang <Jisheng.Zhang at synaptics.com> wrote:

> On Sat, 3 Apr 2021 20:30:53 +0200 (CEST)
> "Maciej W. Rozycki" <macro at orcam.me.uk> wrote:
> 
> > CAUTION: Email originated externally, do not click links or open attachments unless you recognize the sender and know the content is safe.
> > 
> > 
> > On Thu, 1 Apr 2021, Masami Hiramatsu wrote:
> > 
> > > > > > Current riscv's kprobe handlers are run with both preemption and
> > > > > > interrupt enabled, this violates kprobe requirements. Fix this issue
> > > > > > by keeping interrupts disabled for BREAKPOINT exception.  
> > > > >
> > > > > Not only while the breakpoint exception but also until the end of
> > > > > the single step (maybe you are using __BUG_INSN_32 ??) need to be
> > > > > disable interrupts. Can this do that?
> > > > >  
> > > >
> > > > interrupt is disabled during "single step" by kprobes_save_local_irqflag()
> > > > and kprobes_restore_local_irqflag(). The code flow looks like:
> > > >
> > > > do_trap_break()   // for bp
> > > >   kprobe_breakpoint_handler()
> > > >     setup_singlestep()
> > > >       kprobes_restore_local_irqflag()
> > > >
> > > > do_trap_break()  // for ss
> > > >   kprobe_single_step_handler()
> > > >     kprobes_restore_local_irqflag()  
> > >
> > > OK, thanks for the confirmation!  
> > 
> >  Is this approach guaranteed to keep interrupt handling latency low enough
> > for the system not to be negatively affected, e.g. for the purpose of NTP
> > timekeeping?
> 
> IMHO, interrupt latency can't be ensured if kprobes is triggered.

Indeed. The latency depends on what the kprobe user handler does. Of course
it must be as minimal as possible... On x86, it is less than a few microseconds.
Thus it may be a jitter on hard realtime system, but not a big issue on
usual system like NTP timekeeping.

Thank you,

-- 
Masami Hiramatsu <mhiramat at kernel.org>