Re: [RFC 5/5] uprobes: add global breakpoints

[Sebastian Andrzej Siewior <bigeasy@linutronix.de>]

On 08/13/2012 03:16 PM, Oleg Nesterov wrote:
> On 08/09, Sebastian Andrzej Siewior wrote:
>>
>> * Oleg Nesterov | 2012-08-08 15:14:57 [+0200]:
>>
>>>> What I miss right now is an interface to tell the user/gdb that there is a
>>>> program that hit a global breakpoint and is waiting for further instructions.
>>>> A "tail -f trace" does not work and may contain also a lot of other
>>>> informations. I've been thinking about a poll()able file which returns pids of
>>>> tasks which are put on hold. Other suggestions?
>>>
>>> Honestly, I am not sure this is that useful...
>>
>> How would you notify gdb that there is a new task that hit a breakpoint?
>> Or learn yourself?
>
> But why do we need this?

Shouldn't we learn somehow that a process hits a breakpoint? The task
was not yet monitored by gdb.

> OK, you do not need to convince me, I try to never argue with
> new features.

If there is a simple mechanism, I would switch to it. Right now I think
about using this "notification mechanism" to auto-exlude the listener
(and its parents) from the list of possible targets. So I don't freeze
the whole system while I have a breakpoint at malloc() in libc.

> However, I certainly dislike TASK_TRACED in uprobe_wait_traced().
> And sleeping in ->handler() is not fair to other consumers.

I added it as the last task in current consumer. I could move it out of
the consumer loop and freeze it after all consumer are handled but then
I lose the filter member (which is currently NULL, I know).

> And I do not think you should modify ptrace_attach() at all.
> gdb/user can wakeup the task after PTRACE_ATTACH itself.

I see. gdb / strace --pid $num" gdb does PTRACE_ATTACH and waits
afterwords in wait() indefinitely for the SIGSTOP which is blocked
since the process is already in TASK_TRACED. This is nice since the
signals are blocked and are delivered once the task is unfrozed.

> Oleg.

Sebastian