Thread switching and stepping bug

Thread switching and stepping bug

[Jonathan Larmour]

It seems GDB (a fairly recent CVS) doesn't do the right thing when a thread
view has been switched and then the system stepped.

I've got a program with a bunch of threads. The default one is thread 3 and
has a function breakme which I set a breakpoint. The other threads run
other stuff.

If I set a breakpoint on breakme, thread 3 hits it. If I manually step off
that breakpoint, switch to e.g. thread 5 then do another step, GDB can't
recognise that it hit a sensible breakpoint, and instead reports a SIGTRAP.

So the commands I'm doing are:

b breakme
c
[ hits breakpoint in thread 3]
step
thread 5
step

I'm surprised no-one has noticed this before, so is my understanding wrong?

I've tried looking in wait_for_inferior() but got lost quickly :-). I think
the problem may be that step_resume_breakpoint is changed when GDB notices
the thread has changed. But I'm no expert.

Jifl
-- 
Red Hat, Rustat House, Clifton Road, Cambridge, UK. Tel: +44 (1223) 271062
Maybe this world is another planet's Hell -Aldous Huxley || Opinions==mine

Re: Thread switching and stepping bug

[David Smith]

Jonathan,

If I understand you correctly, you may have hit a problem I'm familiar with. 
  Basically your target (which you didn't mention) hasn't implemented 
PREPARE_TO_PROCEED which gets called in infrun.c.  I submitted a patch back 
at the end of March to implement a generic PREPARE_TO_PROCEED which should 
work for most targets.

For more details, here's the start of the thread where we discussed this:

< http://sources.redhat.com/ml/gdb-patches/2001-03/msg00541.html >

I hope this helps.

Jonathan Larmour wrote:

> It seems GDB (a fairly recent CVS) doesn't do the right thing when a thread
> view has been switched and then the system stepped.
> 
> I've got a program with a bunch of threads. The default one is thread 3 and
> has a function breakme which I set a breakpoint. The other threads run
> other stuff.
> 
> If I set a breakpoint on breakme, thread 3 hits it. If I manually step off
> that breakpoint, switch to e.g. thread 5 then do another step, GDB can't
> recognise that it hit a sensible breakpoint, and instead reports a SIGTRAP.
> 
> So the commands I'm doing are:
> 
> b breakme
> c
> [ hits breakpoint in thread 3]
> step
> thread 5
> step
> 
> I'm surprised no-one has noticed this before, so is my understanding wrong?
> 
> I've tried looking in wait_for_inferior() but got lost quickly :-). I think
> the problem may be that step_resume_breakpoint is changed when GDB notices
> the thread has changed. But I'm no expert.
> 
> Jifl



-- 
David Smith
dsmith@redhat.com
Red Hat, Inc.
http://www.redhat.com
256.704.9222 (direct)
256.837.3839 (fax)

Re: Thread switching and stepping bug

[Jonathan Larmour]

David Smith wrote:
> 
> Jonathan,
> 
> If I understand you correctly, you may have hit a problem I'm familiar with.
>   Basically your target (which you didn't mention)

I tried it on v850 and ARM and both failed in the same way so I thought it
was a generic failing.

> hasn't implemented
> PREPARE_TO_PROCEED which gets called in infrun.c.  I submitted a patch back
> at the end of March to implement a generic PREPARE_TO_PROCEED which should
> work for most targets.
> 
> For more details, here's the start of the thread where we discussed this:
> 
> < http://sources.redhat.com/ml/gdb-patches/2001-03/msg00541.html >

Since it's checked in I'll try using the latest CVS to see what happens. My
current GDB's predate your patch.

Jifl
-- 
Red Hat, Rustat House, Clifton Road, Cambridge, UK. Tel: +44 (1223) 271062
Maybe this world is another planet's Hell -Aldous Huxley || Opinions==mine

Re: Thread switching and stepping bug

[Jonathan Larmour]

David Smith wrote:
> 
> Jonathan,
> 
> If I understand you correctly, you may have hit a problem I'm familiar with.
>   Basically your target (which you didn't mention) hasn't implemented
> PREPARE_TO_PROCEED which gets called in infrun.c.  I submitted a patch back
> at the end of March to implement a generic PREPARE_TO_PROCEED which should
> work for most targets.
> 
> For more details, here's the start of the thread where we discussed this:
> 
> < http://sources.redhat.com/ml/gdb-patches/2001-03/msg00541.html >

I've gone through it again with current CVS GDB, and I'm afraid there's no
change in behaviour. I was slightly wrong in my description below actually.
All that's needed is actually:

b breakme
c
[ hits breakpoint in thread 3 ]
thread 5
step

I've already tried to debug it but got lost in wait_for_inferior(). Perhaps
someone could give me an idea what should be the correct course of events
in wait_for_inferior() in the above situation after the final step, and I
can compare with what happens?

Thanks,

Jifl


> Jonathan Larmour wrote:
> 
> > It seems GDB (a fairly recent CVS) doesn't do the right thing when a thread
> > view has been switched and then the system stepped.
> >
> > I've got a program with a bunch of threads. The default one is thread 3 and
> > has a function breakme which I set a breakpoint. The other threads run
> > other stuff.
> >
> > If I set a breakpoint on breakme, thread 3 hits it. If I manually step off
> > that breakpoint, switch to e.g. thread 5 then do another step, GDB can't
> > recognise that it hit a sensible breakpoint, and instead reports a SIGTRAP.
> >
> > So the commands I'm doing are:
> >
> > b breakme
> > c
> > [ hits breakpoint in thread 3]
> > step
> > thread 5
> > step
> >
> > I'm surprised no-one has noticed this before, so is my understanding wrong?
> >
> > I've tried looking in wait_for_inferior() but got lost quickly :-). I think
> > the problem may be that step_resume_breakpoint is changed when GDB notices
> > the thread has changed. But I'm no expert.
> >
> > Jifl
> 
> --
> David Smith
> dsmith@redhat.com
> Red Hat, Inc.
> http://www.redhat.com
> 256.704.9222 (direct)
> 256.837.3839 (fax)

-- 
Red Hat, Rustat House, Clifton Road, Cambridge, UK. Tel: +44 (1223) 271062
Maybe this world is another planet's Hell -Aldous Huxley || Opinions==mine

Re: Thread switching and stepping bug

[David Smith]

Jonathan,

Note that my patch provides the generic PREPARE_TO_PROCEED functionality. 
You'll have to change your tdep file to use generic_prepare_to_proceed().  If 
your target is already multi-arched, it is a one line change in your 
{target}_gdbarch_init() function.  It would look like this:

   set_gdbarch_prepare_to_proceed (gdbarch, generic_prepare_to_proceed);

Give that a shot.

Jonathan Larmour wrote:

> David Smith wrote:
> 
>> Jonathan,
>> 
>> If I understand you correctly, you may have hit a problem I'm familiar with.
>>   Basically your target (which you didn't mention) hasn't implemented
>> PREPARE_TO_PROCEED which gets called in infrun.c.  I submitted a patch back
>> at the end of March to implement a generic PREPARE_TO_PROCEED which should
>> work for most targets.
>> 
>> For more details, here's the start of the thread where we discussed this:
>> 
>> < http://sources.redhat.com/ml/gdb-patches/2001-03/msg00541.html >
> 
> 
> I've gone through it again with current CVS GDB, and I'm afraid there's no
> change in behaviour. I was slightly wrong in my description below actually.
> All that's needed is actually:
> 
> b breakme
> c
> [ hits breakpoint in thread 3 ]
> thread 5
> step
> 
> I've already tried to debug it but got lost in wait_for_inferior(). Perhaps
> someone could give me an idea what should be the correct course of events
> in wait_for_inferior() in the above situation after the final step, and I
> can compare with what happens?
> 
> Thanks,
> 
> Jifl
> 
> 
> 
>> Jonathan Larmour wrote:
>> 
>> 
>>> It seems GDB (a fairly recent CVS) doesn't do the right thing when a thread
>>> view has been switched and then the system stepped.
>>> 
>>> I've got a program with a bunch of threads. The default one is thread 3 and
>>> has a function breakme which I set a breakpoint. The other threads run
>>> other stuff.
>>> 
>>> If I set a breakpoint on breakme, thread 3 hits it. If I manually step off
>>> that breakpoint, switch to e.g. thread 5 then do another step, GDB can't
>>> recognise that it hit a sensible breakpoint, and instead reports a SIGTRAP.
>>> 
>>> So the commands I'm doing are:
>>> 
>>> b breakme
>>> c
>>> [ hits breakpoint in thread 3]
>>> step
>>> thread 5
>>> step
>>> 
>>> I'm surprised no-one has noticed this before, so is my understanding wrong?
>>> 
>>> I've tried looking in wait_for_inferior() but got lost quickly :-). I think
>>> the problem may be that step_resume_breakpoint is changed when GDB notices
>>> the thread has changed. But I'm no expert.
>>> 
>>> Jifl
>> 
>> --
>> David Smith
>> dsmith@redhat.com
>> Red Hat, Inc.
>> http://www.redhat.com
>> 256.704.9222 (direct)
>> 256.837.3839 (fax)



-- 
David Smith
dsmith@redhat.com
Red Hat, Inc.
http://www.redhat.com
256.704.9222 (direct)
256.837.3839 (fax)

Re: Thread switching and stepping bug

[Andrew Cagney]

> Jonathan,
> 
> Note that my patch provides the generic PREPARE_TO_PROCEED functionality. You'll have to change your tdep file to use generic_prepare_to_proceed().  If your target is already multi-arched, it is a one line change in your {target}_gdbarch_init() function.  It would look like this:
> 
>   set_gdbarch_prepare_to_proceed (gdbarch, generic_prepare_to_proceed);
> 
> Give that a shot.

Perhaphs the generic version should be enabled by default?

While this is definitly a departure from the norm - retain default 
behavour even when broken - I suspect it may still be better since 
enabling this fixes a bug.

	Andrew

Re: Thread switching and stepping bug

[David Smith]

Andrew,

I'm a bit nervous about making it the default.  Although I believe that it 
would certainly help 99% of targets, I worry about the last 1%.

Would there be some way of just turning this on for targets that support threads?

Andrew Cagney wrote:

>> Jonathan,
>> 
>> Note that my patch provides the generic PREPARE_TO_PROCEED 
>> functionality. You'll have to change your tdep file to use 
>> generic_prepare_to_proceed().  If your target is already multi-arched, 
>> it is a one line change in your {target}_gdbarch_init() function.  It 
>> would look like this:
>> 
>>   set_gdbarch_prepare_to_proceed (gdbarch, generic_prepare_to_proceed);
>> 
>> Give that a shot.
> 
> 
> Perhaphs the generic version should be enabled by default?
> 
> While this is definitly a departure from the norm - retain default 
> behavour even when broken - I suspect it may still be better since 
> enabling this fixes a bug.
> 
>     Andrew



-- 
David Smith
dsmith@redhat.com
Red Hat, Inc.
http://www.redhat.com
256.704.9222 (direct)
256.837.3839 (fax)

Re: Thread switching and stepping bug

[Andrew Cagney]

> Andrew,
> 
> I'm a bit nervous about making it the default.  Although I believe that it would certainly help 99% of targets, I worry about the last 1%.
> 
> Would there be some way of just turning this on for targets that support threads?

Hmm, that is a good point.  No that isn't currently possible.  Lets 
leave it be.

	Andrew