This problem is about the scheduling policy. I decided not to touch the process search policy (i.e. scanning the list of RUNNING processes) and to change the goodness() that is the score that a process gets to run.

To take into account also the global amount of CPU used by a user i decided to use the struct user_struct * user field of the task_struct.

This is declared in kernel/fork.c and is explicitly said (include/linux/kernel/sched.h)

I add a new field analogous to the ticks counter of the task_struct, and a pointer to link all the user_struct together in a list (used for traversing them).

I also define the root for the list of user_struct's.

The "priority" for this counter is arbitrarily set to 50 for user tasks, and to 100 for tasks without users. I do not know if these values are appropriate. I just picked some value below 1000 (real-time processes goodness) and above zero.

I checked that the kernel boots and can do some tasks.

I did not do any extensive varification that there is no problem. Nor i did any performance evaluation. To this extent i can suggest

1. to start many computation intensive task under a user (say a few fibonacci evaluation with reasonably high value). Then to start a single similar evaluation for another user. This latter should be favoured against the formers because those compete with one another.
2. time a computationally intensive process, when run as single task of a user with the modified kernel and with the original one.

kernel/fork.c
kernel/sched.c

Next   Prev   Up