elfutls using assert

elfutls using assert

[Andrew Cagney]

Roland,

Mike and I noticed that elfutils uses assert() which, in turn, calls 
abort().  Is there a way to override this, having the code instead call 
a callback?  That way, should an assertion fail, frysk will be able to 
at least report the issue before exiting.

Andrew

Re: elfutls using assert

[Roland McGrath]

No, there is no way not to call abort.  It is used for situations that are
pathological in the same way as a clobbered pointer that might lead to a
crash.  If you want to improve recovery from such bugs, you might as well
catch all unexpected fatal signals and diagnose gracefully there (maybe if
things aren't too bad you can throw an exception from the signal handler
and never consult the clobbered data structures while reporting and exiting).

Re: elfutls using assert

[Andrew Cagney]

Unless I'm mistaken, it isn't possible to attach the abort signal as it 
may be sent to any thread?  Anyway, to make the problem more concrete, 
when a panic like this occurs, frysk needs time to pull all inserted 
breakpoints before detaching from the program and exiting.

Andrew

Roland McGrath wrote:
> No, there is no way not to call abort.  It is used for situations that are
> pathological in the same way as a clobbered pointer that might lead to a
> crash.  If you want to improve recovery from such bugs, you might as well
> catch all unexpected fatal signals and diagnose gracefully there (maybe if
> things aren't too bad you can throw an exception from the signal handler
> and never consult the clobbered data structures while reporting and exiting).
>   

Re: elfutls using assert

[Roland McGrath]

> Unless I'm mistaken, it isn't possible to attach the abort signal as it 
> may be sent to any thread?  

You are mistaken.  If you are catching the signal, you will catch all such
signals in whatever thread they are delivered.  Anyway, abort is equivalent
to raise (SIGABRT), which means synchronously to the calling thread (the
signal will appear to happen immediately after the syscall instruction for
the syscall that performs the "raise").

> Anyway, to make the problem more concrete, when a panic like this occurs,
> frysk needs time to pull all inserted breakpoints before detaching from
> the program and exiting.

Like I said, when one of these hits, you are lucky if all memory is not
randomly clobbered already.  But if you just avoid anything that would call
into elfutils libs at all, you might be avoiding the memory that is
definitely clobbered already, and be lucky enough to survive long enough to
withdraw gracefully.


Thanks,
Roland

Re: elfutls using assert

[Andrew Cagney]

Roland McGrath wrote:
>> Unless I'm mistaken, it isn't possible to attach the abort signal as it 
>> may be sent to any thread?  
>>     
>
>   

Good, I'll add code to attempt that.

> You are mistaken.  If you are catching the signal, you will catch all such
> signals in whatever thread they are delivered.  Anyway, abort is equivalent
> to raise (SIGABRT), which means synchronously to the calling thread (the
> signal will appear to happen immediately after the syscall instruction for
> the syscall that performs the "raise").
>
>   
>> Anyway, to make the problem more concrete, when a panic like this occurs,
>> frysk needs time to pull all inserted breakpoints before detaching from
>> the program and exiting.
>>     
>
> Like I said, when one of these hits, you are lucky if all memory is not
> randomly clobbered already.  But if you just avoid anything that would call
> into elfutils libs at all, you might be avoiding the memory that is
> definitely clobbered already, and be lucky enough to survive long enough to
> withdraw gracefully.
>
>
> Thanks,
> Roland
>