Re: [PATCH 2/2] gdb: c++ify btrace_target_info

[Simon Marchi <simon.marchi@polymtl.ca>]

On 9/7/23 06:42, Metzger, Markus T wrote:
> Hello Simon,
> 
> Thanks for your review.
> 
>>>    if (errcode == BTRACE_ERR_NONE)
>>> -    xfree (tinfo);
>>> +    delete gtinfo;
>>
>> A question about this, even if it's not introduced by this patch... if
>> errcode is not BTRACE_ERR_NONE, who deletes gtinfo?
> 
> It would be leaked.  Since tracing couldn't be disabled, it would still
> be in use, so leaking it seems better than deleting an in-use object
> and risking either crash or data corruption.

I'm not sure I follow.  Even if something fails when trying to disable
btrace (like a call to the OS fails for an unknown reason), GDB can
still decide it doesn't want to do tracing / consume btrace data and get
rid of the object.

> Now, that cannot happen today, since both linux_disable_pt and
> linux_disable_bts return BTRACE_ERR_NONE.

Ok, well if these operations can't fail I think they should not return
an error code (for the same reason I wrote below about hypothetical
scenarios).

>>> @@ -100,6 +101,15 @@ struct btrace_target_info
>>>      struct btrace_tinfo_pt pt;
>>>    } variant;
>>>  #endif /* HAVE_LINUX_PERF_EVENT_H */
>>> +
>>> +
>>> +  linux_btrace_target_info (ptid_t _ptid, const struct btrace_config &_conf)
>>> +    : ptid (_ptid), conf (_conf), variant ({})
>>> +    {}
>>> +
>>> +  linux_btrace_target_info (ptid_t _ptid)
>>> +    : ptid (_ptid), conf ({}), variant ({})
>>
>> It's probably not needed to explicitly initialize conf and variant.
> 
> That's an interesting question.  My interpretation is that not specifying a
> ctor-initializer would leave things undefined*, whereas supplying an empty
> aggregate would zero-initialize everything.
> 
> *we'd default-construct conf, which has an implicitly defined default
> constructor, which then default-constructs all the data members, and so on,
> until we reach plain integers, on which no initialization is performed.

Ok, replied to that in my response to v2.

>>>  {
>>>    /* The ptid of the traced thread.  */
>>>    ptid_t ptid;
>>>
>>>    /* The obtained branch trace configuration.  */
>>>    struct btrace_config conf;
>>
>> So, these two fields are common to both btrace_target_info sub-classes.
>> Do you intend to move them to the base class?  That would allow getting
>> rid of some (all?) casts in remote.c.
> 
> No.  I think it's better to leave the base class empty to give full freedom
> to targets.

Do you have an actual case of btrace target that wouldn't use these
fields? My stance on it is: this is all internal GDB code, it's not an
API we expose.  I don't think we should complicate things for
hypothetical scenarios, when we always have the freedom to re-work
things later if need be.  I think it makes more sense to keep things
simple for what we have today.

Simon