Andi,
I can provide some details about a more specific use-case that arises from
the java language and runtime environment.
In the java world, a given thread can be renamed at any time by any other
thread, by calling a method on the receiver object that represents the
thread. This will give the thread a new name in java, and we'd like to
make this work with Linux thread naming as well, so that the java thread
names are visible from the various O/S utilities such as ps.
Using the existing prctl mechanism, java threads can assign their own
names to themselves when they are started. However, more generally, at
any other time, in order to make the existing java thread name method call
(callable by any thread in the java virtual machine) do the additional
work of renaming the same thread in Linux, the process requires several
steps:
1. The naming thread must asynchronously provide a notification to the
named thread that a name change has occurred, and must store the name in a
shared location
2. The named thread must be granted CPU time by the Linux scheduler
3. The named thread must poll for the notication, and must apply any
assigned name change prior to doing what it would normally be doing
Some undesirable drawbacks of this include the async communication, the
delay waiting for the named thread to get CPU time for the name change to
occur, the polling activity in the named thread, and the delay in the
named thread changing its own name prior to resuming its normal activity
Using the Java API, thread name changes can be completed without the
involvement of the thread being named. As John has described, java
programmers can choose to have a single thread assign names to all others.
Programmers may choose to rename threads even when the threads being
named are not being granted any CPU time. Programmers may choose to
rename threads in an ad-hoc manner. We'd like to extend this naming
activity to the Linux naming functionality without the drawbacks listed
above.
Regards,
Sean Foley
J9 Real-Time Java Ottawa Technical Lead
[email protected] (613)356-5012
john stultz <[email protected]> wrote on 10/26/2009 07:56:20 PM:
> On Sat, 2009-10-24 at 05:54 +0200, Andi Kleen wrote:
> > On Wed, Oct 21, 2009 at 07:48:22PM -0700, Arjan van de Ven wrote:
> > > On Thu, 22 Oct 2009 02:49:19 +0200
> > > Andi Kleen <[email protected]> wrote:
> > >
> > > > > you're pretty much the first now who touches ->comm from
> > > > > not-the-thread-itself.... are you sure that is safe?
> > > >
> > > > It's not, there is no locking. On the other hand nothing should
crash,
> > > > just users might see half rewritten data.
> > >
> > > .. with strings this is tricky though.. if the new string is longer
> > > than the old one the terminating zero might just be missed etc.
> >
> > Good point.
> > He needs to first set the last byte to zero to avoid that.
> > But better probably to do a proper lock on all readers. Or not add
> > this feature at all (does it have a strong use case?)
>
> Thread naming is really helpful for debugging a large multi-threaded
> application. But currently it requires the threads to name themselves.
>
> In some applications there may be a dispatch thread that spawns off the
> siblings, and it has more context for naming the threads then the
> threads do themselves.
>
> So in that case, currently in order to have named threads, the
> application's dispatch thread assigns names, and the threads have to
> occasionally poll to see if they need to name themselves to something.
>
> So they can get it to work, but its ugly and a big pain. Other OSes
> support pthread_set_name_np() functionality, which makes things easy for
> them, so they've requested to see if Linux can't do something similar.
>
> thanks
> -john
>