scx/BREAKING_CHANGES.md

Breaking Changes

sched_ext is still experimental and while we're trying to reduce the number of breaking changes, sometimes they're neccessary to fix bugs or improve the usability.

Below are the changes in both the sched_ext kernel tree and the associated commits in this repo.

scx: Rename prep_enable() and cancel_enable(), add exit_task() [Kernel commit / PR] [scx commit]

• ops.prep_enable() is now called ops.init_task()
  • struct scx_enable_args is now struct scx_init_task_args
• ops.enable() and ops.disable() are now called when a task enters and exits sched_ext respectively. This is the same for when all tasks are using scx, but the callbacks can now be fired multiple times for tasks switching their sched policy.
  • No longer passing struct scx_enable_args *args to ops.enable().
• ops.cancel_enable() has been removed, and ops.exit_task() (explained below) is invoked in its stead.
• ops.exit_task() has been added, and is called exactly once when a task is exiting if ops.init_task() had been successfully invoked on the task previously (or would have been invoked if the callback was defined).
  • Called in lieu of ops.cancel_enable() as described above
  • Called with struct scx_exit_task_args *args instead of scx_enable_args *args.

---

scx: Add scx_bpf_select_cpu_dfl() kfunc [Kernel commit…Kernel commit, PR] [scx commit]

• The default CPU selection logic operates differently now. We no longer pass SCX_ENQ_LOCAL when the default CPU selection has found a core to schedule. Callers can instead use scx_bpf_select_cpu_dfl() to get the same behavior and then decide whether to direct dispatch or not.
• Tasks can now be direct-dispatched from ops.select_cpu().