Re: kgdb for mainline kernel: core-lite [patch 1/3]

["Amit S. Kale" <amitkale@emsyssoft.com>]

On Monday 08 Mar 2004 8:49 pm, Tom Rini wrote:
> On Mon, Mar 08, 2004 at 05:14:05PM +0530, Amit S. Kale wrote:
> > On Monday 08 Mar 2004 4:50 pm, Amit S. Kale wrote:
> > > On Monday 08 Mar 2004 4:37 pm, Andrew Morton wrote:
> > > > "Amit S. Kale" <amitkale@emsyssoft.com> wrote:
> > > > > On Monday 08 Mar 2004 3:56 pm, Andrew Morton wrote:
> > > > >  > "Amit S. Kale" <amitkale@emsyssoft.com> wrote:
> > > > >  > > Here are features that are present only in full kgdb:
> > > > >  > >  1. Thread support  (aka info threads)
> > > > >  >
> > > > >  > argh, disaster.  I discussed this with Tom a week or so ago when
> > > > >  > it looked like this it was being chopped out and I recall being
> > > > >  > told that the discussion was referring to something else.
> > > > >  >
> > > > >  > Ho-hum, sorry.  Can we please put this back in?
> > > > >
> > > > >  Err., well this is one of the particularly dirty parts of kgdb.
> > > > > That's why it's been kept away. It takes care of correct thread
> > > > > backtraces in some rare cases.
> > > >
> > > > Let me just make sure we're taking about the same thing here.  Are
> > > > you saying that with kgdb-lite, `info threads' is completely missing,
> > > > or does it just not work correctly with threads (as opposed to
> > > > heavyweight processes)?
> > >
> > > info threads shows a list of threads. Heavy/light weight processes
> > > doesn't matter. Thread frame shown is incorrect.
> > >
> > > I looked at i386 dependent code again. Following code in it is
> > > incorrect. I never noticed it because this code is rarely used in full
> > > version of kgdb:
> > >
> > > +void sleeping_thread_to_gdb_regs(unsigned long *gdb_regs, struct
> > > task_struct *p)
> > > ....
> > > +	gdb_regs[_EBP] = *(int *)p->thread.esp;
> > >
> > > We can't guss ebp this way. This line should be removed.
> > >
> > > +	gdb_regs[_DS] = __KERNEL_DS;
> > > +	gdb_regs[_ES] = __KERNEL_DS;
> > > +	gdb_regs[_PS] = 0;
> > > +	gdb_regs[_CS] = __KERNEL_CS;
> > > +	gdb_regs[_PC] = p->thread.eip;
> > > +	gdb_regs[_ESP] = p->thread.esp;
> > >
> > > This should be gdb_regs[_ESP] = &p->thread.esp
> >
> > That's not correct it. It should be gdb_regs[_ESP] = p->thread.esp;
> > Even with these changes I can't get thread listing correctly.
> >
> > Here is the intrusive piece of code that helps get thread state
> > correctly. Any ideas on cleaning it?
>
> Here's where what Andi said about being able to get the pt_regs stuff
> off the stack (I think that's what he said at least) started to confuse
> me slightly.  But if I understand it right (and I never got around to

That's an in interesting idea. We can at least get first pt_regs from the 
stack. That way we can get to user space eip in case a thread is in user 
space.

The do_schedule and save all regs is definitely required if we want _exact_ 
state of all kernel threads. [exact state: all registers]

Finding function frames below pt_regs becomes difficult or impossible, as 
exact dept of stack for each function in a stack is not known. EBP chains 
allow walking "up" but not "down".
-Amit


> testing this) we can replace the do_schedule() stuffs at least with
> something like kgdb_get_pt_regs(), since (and I lost my notes on this,
> so it's probably not quite right) (thread_info->esp0)-1