compat.bpf.h: Add override for scx_bpf_dsq_peek version gate

[bboymimi]

a3f2784 ("compat.bpf.h: Gate scx_bpf_dsq_peek kfunc behind kernel
version 7.1.0") gates the lockless peek kfunc by LINUX_KERNEL_VERSION,
but the version check misses stable backports. A kernel that has both
upstream FIFO fixes

71d7847cad44 ("sched_ext: Fix scx_bpf_dsq_peek() with FIFO DSQs")
2f2ea7709266 ("sched_ext: Use dsq->first_task instead of list_empty()
in dispatch_enqueue() FIFO-tail")

cherry-picked but still reporting < 7.1.0 ends up on the iterator
fallback. CachyOS users on 6.18.31 reported a 4x FPS drop in Cyberpunk
2077.

Add __COMPAT_scx_allow_lockless_peek_unsafe as a __weak rodata bool.
When true, bpf_ksym_exists() alone is enough to take the fast path.
Expose it in scx_lavd as --allow-lockless-peek-unsafe. Default false,
so kernels without the fixes stay on the iterator path. Only enable
on kernels that have both fixes backported.

[daidavid]

I'm not too convinced that using non-peek can cause this level of regression or else we'd have very poor performance for older kernels and/or older LAVD versions. Do we understand what causes the perf drop and can we replicate the regression in other microbenchmarks?

[multics69]

Let's wait until we collect more evidence of the performance regression.

[ms178]

Thanks a lot for the patch which I've just tested with the latest scx-scheds-git main branch on Kernel 6.18.32 in these two scenarios:

Scenario A: LAVD without --allow-lockless-peek-unsafe --> this yielded the previously reported massive performance regression and massive stuttering

Possibly related dmesg output:

[19. Mai 10:07] NOHZ tick-stop error: local softirq work is pending, handler #80!!!
[  +0,000000] NOHZ tick-stop error: local softirq work is pending, handler #200!!!
[  +0,000554] NOHZ tick-stop error: local softirq work is pending, handler #200!!!

Scenario B: LAVD with --allow-lockless-peek-unsafe --> this brought back the expected performance level

[multics69]

Thanks a lot for the patch which I've just tested with the latest scx-scheds-git main branch on Kernel 6.18.32 in these two scenarios:

Scenario A: LAVD without --allow-lockless-peek-unsafe --> this yielded the previously reported massive performance regression and massive stuttering

Possibly related dmesg output:

[19. Mai 10:07] NOHZ tick-stop error: local softirq work is pending, handler #80!!!
[  +0,000000] NOHZ tick-stop error: local softirq work is pending, handler #200!!!
[  +0,000554] NOHZ tick-stop error: local softirq work is pending, handler #200!!!

Scenario B: LAVD with --allow-lockless-peek-unsafe --> this brought back the expected performance level

What's the hardware you are using? It sounds like a pretty beefy machine with many CPUs?

[ms178]

What's the hardware you are using? It sounds like a pretty beefy machine with many CPUs?

@multics69 As mentioned in #3580 (comment), I use an Intel 14700KF (Raptor Lake) with a Z690 motherboard, 64 GB DDR4-3600, AMD Vega 64. The CPU comes with 8 P-Cores and 12 E-Cores.