[REGRESSION] 6.12.y: d66792919d4f (sched/deadline: Use revised wakeup rule for dl_server) causes latencies up to 50ms with PREEMPT_RT

Hi,

I am reporting a regression which was introduced by d66792919d4f on 6.12.y.
Since this commit, cyclictest reports latencies up to 50 milliseconds, 
on kernels with CONFIG_PREEMPT_RT=y.

Steps to reproduce:
1. run a load (e.g. stress-ng --cpu 4 --io 2 --vm 2 --vm-bytes 128M)
2. run cyclictest (e.g. cyclictest -a -t -m -p 80 -i 250 -d 0)

cyclictest results on the current linux-6.12.y branch (tag v6.12.87):
# /dev/cpu_dma_latency set to 0us
policy: fifo: loadavg: 9.37 9.21 6.90 9/211 978
T: 0 ( 884) P:80 I:250 C:4688252 Min: 3 Act: 6 Avg: 6 Max: 51956
T: 1 ( 885) P:80 I:250 C:4688051 Min: 3 Act: 7 Avg: 6 Max: 50106
T: 2 ( 886) P:80 I:250 C:4688242 Min: 3 Act: 6 Avg: 6 Max: 51965
T: 3 ( 887) P:80 I:250 C:4688434 Min: 3 Act: 12 Avg: 8 Max: 59

cyclictest results on 6.12.y with d66792919d4f reverted:
# /dev/cpu_dma_latency set to 0us
policy: fifo: loadavg: 9.43 9.50 9.44 8/204 5758
T: 0 ( 862) P:80 I:250 C:272329322 Min: 3 Act: 6 Avg: 6 Max: 57
T: 1 ( 863) P:80 I:250 C:272329324 Min: 3 Act: 7 Avg: 6 Max: 77
T: 2 ( 864) P:80 I:250 C:272329322 Min: 3 Act: 7 Avg: 6 Max: 68
T: 3 ( 865) P:80 I:250 C:272329322 Min: 3 Act: 16 Avg: 7 Max: 81

This is reproducible on multiple machines.

It looks like the timer fires and there is also a sched_waking event in 
the trace, but the cyclictest thread does not get scheduled for another 
50ms.

I found this, because Debian updated its rt kernel from 6.12.74 to 6.12.85.
The issue was also present with upstream 6.12.85 and HEAD, but not with 
6.12.74, so I started bisecting and eventually found d66792919d4f.

Is it possible to revert the commit?

I can provide traces or help with testing if needed.

Thanks
Lukas Beckmann

#regzbot introduced: d66792919d4f