sched/eevdf: Fix pick_eevdf()
authorBenjamin Segall <bsegall@google.com>
Sat, 30 Sep 2023 00:09:30 +0000 (17:09 -0700)
committerPeter Zijlstra <peterz@infradead.org>
Mon, 9 Oct 2023 07:48:33 +0000 (09:48 +0200)
commitb01db23d5923a35023540edc4f0c5f019e11ac7d
treed6fd6bedd01a4137f41f4b54cb810022dae1af07
parent8dafa9d0eb1a1550a0f4d462db9354161bc51e0c
sched/eevdf: Fix pick_eevdf()

The old pick_eevdf() could fail to find the actual earliest eligible
deadline when it descended to the right looking for min_deadline, but
it turned out that that min_deadline wasn't actually eligible. In that
case we need to go back and search through any left branches we
skipped looking for the actual best _eligible_ min_deadline.

This is more expensive, but still O(log n), and at worst should only
involve descending two branches of the rbtree.

I've run this through a userspace stress test (thank you
tools/lib/rbtree.c), so hopefully this implementation doesn't miss any
corner cases.

Fixes: 147f3efaa241 ("sched/fair: Implement an EEVDF-like scheduling policy")
Signed-off-by: Ben Segall <bsegall@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/xm261qego72d.fsf_-_@google.com
kernel/sched/fair.c