Re: [patch, gdbserver] Uninsert bpkt when regular and fast tracepoint are set at the same address

[Pedro Alves <pedro@codesourcery.com>]

On Thursday 27 October 2011 19:38:31, Stan Shebs wrote:
> On 10/26/11 7:56 PM, Yao Qi wrote:
> > Hi,
> > I find a program will receive segv fault when I set a regular tracepoint
> > and a fast tracepoint at the same address, start tracing and resume program.
> >
> > gdbserver has taken care of this situation in many places of the code,
> > when uninserting breakpoint or fast tracepoint, write_inferior_memory is
> > called to take care of layering breakpoints on top of fast tracepoints.
> >   However, it is not right to me.  Here is an example to illustrate this
> > problem.
> >
> > Supposing I set a regular tracepoint and a fast tracepoint on 0x080484fc,
> >
> >      0x080484fc<+3>:     e8 f3 ff ff ff  call   0x80484f4<func>
> >
> > During insertion, trap insn (for regular tracepoint) and jmp insn (for
> > fast tracepoint) are inserted, and gdbserver takes care of them to make
> > sure trap insn is *always* inserted on top of jmp insn.
> 
> I'm looking at this and wondering, why are we inserting the fast 
> tracepoint jump insn at all?

I think the way we do things currently is simpler.  Consider:

 1 - ftrace foo (5 bytes)
 2 - tstart, installs fast tracepoint
 3 - b foo, sets breakpoint (would remove the jmp?)
 4 - del breakpoint $bkpt_foo (would re-insert insert the jmp?)

If in 3 the answer is no, then we have to handle the jmp being
inserted anyway.  If it is yes, then, it looks like extra
weird work.  Same for step 4.  This introduces unnecessary
coupling between different "kinds" of breakpoints.

As is, we think in terms of range of addresses we're
writting to / reading from.

> Shouldn't be it sufficient to let the trap handler do the work of both 
> slow and fast tracepoints at that location?  
> Since hitting the trap has 
> already put us on the slow path, there's not going to any noticeable 
> additional penalty for not going to the IPA and interpreting conditional 
> bytecodes instead of compiled ones, etc.  We may not even need to sync 
> trace buffers (I'm not sure about that, code is tricky).

That's actually kind of a separate issue, and we already do that.
gdbserver/tracepoint.c has this comment:

  /* Presently, gdbserver doesn't run compiled conditions, only the
     IPA does.  If the program stops at a fast tracepoint's address
     (e.g., due to a breakpoint, trap tracepoint, or stepping),
     gdbserver preemptively collect the fast tracepoint.  Later, on
     resume, gdbserver steps over the fast tracepoint like it steps
     over breakpoints, so that the IPA doesn't see that fast
     tracepoint.  This avoids double collects of fast tracepoints in
     that stopping scenario.  Having gdbserver itself handle the fast
     tracepoint gives the user a consistent view of when fast or trap
     tracepoints are collected, compared to an alternative where only
     trap tracepoints are collected on stop, and fast tracepoints on
     resume.  When a fast tracepoint is being processed by gdbserver,
     it is always the non-compiled condition expression that is
     used.  */

-- 
Pedro Alves