Re: [PATCHv10 2.6.35-rc6-tip 8/14] trace: Extract out common code for kprobes/uprobes traceevents.

[Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>]

Srikar Dronamraju wrote:
>>> ---
>> [...]
>>> +/* Recursive argument parser */
>>> +static int parse_probe_arg(char *arg, const struct fetch_type *t,
>>> +		     struct fetch_param *f, int is_return, bool is_kprobe)
>> If you use "bool" for "is_kprobe", change "is_return" type too.
> 
> Okay. 
> 
> 
>> And, maybe you missed that the fetch function supports "string" type now,
>> which needs a bit different manner for storing fetched value. You can find
>> store_trace_args() function in trace_kprobe.c.
> 
> Yes, I have seen your changes for supporting string type.  Though all
> the fetch functions are in common code, uprobe based probes for now
> supports register fetching only. We have to add support for other
> types gradually. Please let me know if you see a reason to change the
> way we fetch even when we only support register type access.

Ah, OK. So uprobe-tracer restricts specifying string type on register :)

>> BTW, current fetch functions doesn't support fetching "paged-out" user-variables
>> because kprobe can't sleep inside its handler.
>> However, user-space memory can be paged out, and I assume that uprobes allows
>> its handler to I/O (and yield). If so, it can wait for accessing paged-out
>> variable, can't it?
> 
> Yes, the uprobes handler might sleep and hence we would have to
> handle accessing the paged-out user variables. When perf-uprobes
> starts supporting dwarf based probing, we should look into these
> issues.  Currently its a todo.

Great!

Thank you,

-- 
Masami HIRAMATSU
2nd Research Dept.
Hitachi, Ltd., Systems Development Laboratory
E-mail: masami.hiramatsu.pt@hitachi.com