`gpu-init job-release` releases lock to early

[donald]

mxq/helper/gpu-setup

Line 271 in f3d9fb8

rm $d/pid

We should keep the lock while we modify the access file, otherwise we race with a new allocation.

[thomas]

Wow, this was fast !

So removal must be done at a different location in the script, or will it be kind of an extra/cleanup task?

[donald]

This is not the answer to #139!

I just spotted this. Its a race. I think the bad outcome is more theoretically, because it is near to impossible to occur in reality. The result would be, that the cuda access files are owned by root and not be the uid of the job, so that accessing the gpus would fail.

   MXQ           job1                        job2
* fork job1
                 * other initialization
                 * reserve gpu:
                 * * find slot without pid
                 * * change access to UID
                 * run user program
                 * exit
* fork job2                           
                                             *  other initialization       
* cleanup job 1:
* * rm .../pid
                                             * reserve gpu:
                                             * * find slot without pid
                                             * * change access to UID
* * change access to root
                                             * run user program

So now job2 would have no access to its gpu.