Re: Using getpid() often, another way? [was Re: clone() <-> getpid() bug in 2.6?]

["Dominik Straßer" <einmal_rupert@gmx.de>]

Linus Torvalds wrote:

>On Sun, 6 Jun 2004, Russell Leighton wrote:
>  
>
>>Linus said he could not imagine a program using getpid() more than a 
>>handful of times...
>>    
>>
>
>No, I said that I could not imagine using it more than a handful of times 
>_except_ for the cases of:
>
> - thread identification without a native thread area
> - benchmarking.
>
>(and in both of these cases it is _wrong_ to cache the pid value)
>
>  
>
>>well, (I am ashamed to admit it) I found this getpid() bug with just 
>>such a program.
>>
>>Could someone can suggest an alternative to using getpid() for me? 
>>Here's the problem...
>>
>>I have a library that creates 2 threads using clone().
>>    
>>
>
>Your problem falls under the thread ID thing. It's fine and understandable 
>to use getpid() for that, although you could probably do it faster if you 
>are willing to use the support that modern kernels give you and that glibc 
>uses: the "TLS" aka Thread Local Storage support.
>
>Thread-local storage involved having a user-mode register that points some 
>way to a special part of the address space. On x86, where the general 
>register set is very limited and stealing a general reg would thus be bad, 
>it uses a segment and loads the TLS pointer into a segment register 
>(segment registers are registers too - and since nobody sane uses them for 
>anything else these days, both %gs and %fs are freely usable).
>
>  
>
>> Would gettid() be any better?
>>    
>>
>
>You'd avoid this particular glibc bug with gettid.
>  
>
I am facing the following problem:
I want to sum up the time spent in the main thread + all threads that 
ever existed.
getrusage dosn't work (and didn't do so in pre-NPTL-times) as the time 
spent in threads is not taken into account.
To work around this problem I created a map pid->time used which used 
getpid in the pre-NPTL-time and looked up the time in /proc/<pid>. As 
this doesn't work with NPTL, changed it to use the gettid syscall as I 
didn't find a saner way.

Regards

Dominik