Re: [PATCH v2 sched_ext/for-7.2 0/2] sched_ext: Add cookie API for early qseq capture

[Tejun Heo <tj@kernel.org>]

Hello, Cheng-Yang.

The cookie/qseq vocabulary leaks the implementation. The API is
really a dispatch transaction where commit fails if the task is no
longer ours to dispatch (dequeued, or claimed by someone else).
Naming along those lines would describe what the API does, e.g.:

  scx_bpf_dsq_insert_begin(p)              -> u64 cookie
  scx_bpf_dsq_insert_commit(p, dsq_id, enq_flags, cookie)

The internal scx_dsq_insert_commit() helper would need a rename.
The kdoc can then describe the transaction directly without
referring to qseq or ops_state.

For the selftest: the BPF program calls insert from ops.enqueue(),
which goes through mark_direct_dispatch() and never stages an entry
in the dispatch buffer. The cookie is silently ignored. The test
would still pass if patch 1/2 stored 0 in place of the cookie. To
actually exercise the race-detection window, the insert needs to
happen from ops.dispatch() with something invalidating the task in
between (e.g. an affinity flip).

Thanks.

--
tejun