Re: [PATCH v3 05/11] arm64: Add call_break_hook() to early_brk64() for early kgdb

[Doug Anderson]

Hi,

On Tue, May 12, 2020 at 11:17 PM Will Deacon <will@xxxxxxxxxx> wrote:
>
> Hey Doug,
>
> On Tue, May 12, 2020 at 08:27:50AM -0700, Doug Anderson wrote:
> > On Tue, May 12, 2020 at 12:36 AM Will Deacon <will@xxxxxxxxxx> wrote:
> > > On Mon, May 11, 2020 at 03:45:02PM -0700, Doug Anderson wrote:
> > > > On Mon, May 11, 2020 at 7:59 AM Will Deacon <will@xxxxxxxxxx> wrote:
> > > > > On Tue, Apr 28, 2020 at 02:13:45PM -0700, Douglas Anderson wrote:
> > > > > > diff --git a/arch/arm64/kernel/traps.c b/arch/arm64/kernel/traps.c
> > > > > > index cf402be5c573..a8173f0c1774 100644
> > > > > > --- a/arch/arm64/kernel/traps.c
> > > > > > +++ b/arch/arm64/kernel/traps.c
> > > > > > @@ -1044,6 +1044,9 @@ int __init early_brk64(unsigned long addr, unsigned int esr,
> > > > > >       if ((comment & ~KASAN_BRK_MASK) == KASAN_BRK_IMM)
> > > > > >               return kasan_handler(regs, esr) != DBG_HOOK_HANDLED;
> > > > > >  #endif
> > > > > > +     if (call_break_hook(regs, esr) == DBG_HOOK_HANDLED)
> > > > > > +             return 0;
> > > > >
> > > > > I think this just means we're not running debug_traps_init() early enough,
> > > > > and actually the KASAN early handler is unnecessary too.
> > > > >
> > > > > If we call debug_traps_init() directly from setup_arch() and drop the
> > > > > arch_initcall(), can we then drop early_brk64 entirely?
> > > >
> > > > It seems to work in my testing.  ...but the worry I have is the
> > > > comment right before trap_init().  It says:
> > > >
> > > > /* This registration must happen early, before debug_traps_init(). */
> > >
> > > I /think/ the reason for this is because debug_traps_init() replaces the
> > > BRK vector, so if that runs before the break hooks have been registered
> > > for e.g. BUG() then BUG() won't work during that window. Hmm, so dropping
> > > early_brk64 is problematic after all. Damn.
> > >
> > > Is trap_init() early enough for you? If so, we could call debug_traps_init()
> > > from traps_init() after registering the break hooks.
> >
> > "Early enough" is a subjective term, of course.  The earlier we can
> > init, the earlier we can drop into the debugger.  ...but, of course,
> > everyone thinks their feature is the most important and should be
> > first, so let's see...
> >
> > Certainly if we waited until trap_init() it wouldn't be early enough
> > to set "ARCH_HAS_EARLY_DEBUG".  Setting that means that debugging is
> > ready when early params are parsed and those happen at the start of
> > setup_arch().  The call to trap_init() happens a bit later.
> >
> > If we decide that we just don't care about getting
> > "ARCH_HAS_EARLY_DEBUG" to work then the earliest we'll be able to
> > break into the debugger (via kgdbwait) is dbg_late_init().  That
> > _does_ happen after trap_init() so your solution would work.
> >
> > As a person who spends most of his time in driver land, it wouldn't be
> > the end of the world to wait for dbg_late_init().  That's still much
> > earlier than most code I'd ever debug.  ...and, bonus points is that
> > if we hit a crash any time after earlyparams we _will_ still drop into
> > the debugger.  It's only breakpoints that won't be available until
> > dbg_late_init().
> >
> >
> > tl;dr:
> >
> > * If we care about "kgdbwait" and breakpoints working as early as
> > possible then we need my patch.
> >
> > * If we are OK w/ a slightly later "kgdbwait" then I think we can move
> > debug_traps_init() to trap_init() and get rid of the early version.
> >
> >
> > Please let me know which way you'd like to proceed.
>
> Let's go with the trap_init() approach for now, and we can revisit it later
> if somebody has a compelling reason to initialise things earlier. However,
> I don't think you can remove early_brk64(), as it's needed for BUG() to
> work correctly.

Posted at:

20200513160501.1.I0b5edf030cc6ebef6ab4829f8867cdaea42485d8@changeid">https://lore.kernel.org/r/20200513160501.1.I0b5edf030cc6ebef6ab4829f8867cdaea42485d8@changeid

I'll also reply to the v4 version of this patch to point at it.

-Doug