scx_layered: Add topology aware preemption

[hodgesds]

Add topology aware preemption, which changes the order of preemption to first attempt to preempt CPUs in the local LLC, local NUMA node and finally across all NUMA nodes. This could probably be made a little more efficient in the use of the bpf cpumasks, but should work for now. The old logic is still available behind the -t (disable topology) flag. Resolves #659

[hodgesds]

Example stats output:

###### Thu, 19 Sep 2024 12:20:31 -0400 ######
tot=    776 local=17.53 open_idle=10.05 affn_viol=27.71 proc=5ms
busy= 86.0 util= 2645.9 load=   3806.8 fallback_cpu= 12
excl_coll=0.00 excl_preempt=0.00 excl_idle=0.00 excl_wakeup=0.00
  chrome: util/frac=    3.3/  0.1 load/frac=      6.8/  0.2 tasks=   157
          tot=     29 local=20.69 wake/exp/reenq=79.31/ 0.00/ 0.00
          keep/max/busy=37.93/ 0.00/ 0.00 kick= 0.00 yield/ign= 0.00/    0
          open_idle= 6.90 mig=82.76 xnuma_mig= 0.00 xllc_mig=34.48 affn_viol= 0.00
          preempt/first/xllc/xnuma/idle/fail=10.34/10.34/ 0.00/ 0.00/ 6.90/62.07 min_exec=93.10/   9.54ms
          cpus=  2 [  2,  2] 00000003
  daemon: util/frac=    2.6/  0.1 load/frac=   2188.3/ 57.5 tasks=   148
          tot=    462 local= 1.52 wake/exp/reenq=81.60/16.45/ 0.43
          keep/max/busy= 0.00/ 0.00/ 6.28 kick= 0.00 yield/ign=12.77/    0
          open_idle= 0.00 mig=25.32 xnuma_mig= 0.00 xllc_mig= 9.52 affn_viol=46.54
          preempt/first/xllc/xnuma/idle/fail= 0.00/ 0.00/ 0.00/ 0.00/ 0.00/ 0.00 min_exec= 0.00/   0.00ms
          cpus=  2 [  2,  2] 0000000c
  normal: util/frac=    0.0/  0.0 load/frac=      0.0/  0.0 tasks=     0
          tot=      0 local= 0.00 wake/exp/reenq= 0.00/ 0.00/ 0.00
          keep/max/busy= 0.00/ 0.00/ 0.00 kick= 0.00 yield/ign= 0.00/    0
          open_idle= 0.00 mig= 0.00 xnuma_mig= 0.00 xllc_mig= 0.00 affn_viol= 0.00
          preempt/first/xllc/xnuma/idle/fail= 0.00/ 0.00/ 0.00/ 0.00/ 0.00/ 0.00 min_exec= 0.00/   0.00ms
          cpus=  0 [  0,  0] 00000000
  user  : util/frac= 2640.0/ 99.8 load/frac=   1611.7/ 42.3 tasks=    74
          tot=    285 local=43.16 wake/exp/reenq=25.26/27.72/ 3.86
          keep/max/busy=418.9/14.74/ 4.56 kick=11.93 yield/ign= 0.00/    0
          open_idle=26.67 mig=43.16 xnuma_mig= 0.00 xllc_mig=25.26 affn_viol= 0.00
          preempt/first/xllc/xnuma/idle/fail= 3.16/ 0.00/ 1.40/ 0.00/18.95/22.81 min_exec=45.96/   6.15ms
          cpus=  8 [  8, 10] 00000ff0

[JakeHillion]

This looks good to me. I wonder if it's worth changing the stuff in the new if (!disable_topology) into two separate functions? Might make things easier to follow.

[JakeHillion]

Is there a reason we weren't committing this before? Imo it should be committed, just surprising to see it now.

[hodgesds]

I'm not sure.... I think @likewhatevs recently removed the lock files from .gitignore so we could have reproducible builds. So maybe it hasn't been committed yet?

[likewhatevs]

I did remove the git ignore, but it is kinda weird one would be there at all tbh. Like a top-level Cargo.lock I'd get (workspace level), or a project level Cargo.lock I'd also get (project level).

In the middle seems off to me.

[hodgesds]

I think I know what happened, I think I created this branch before the lock files and before some of the refactoring of the builds. Then after rebasing the old lock file was still around and I added it on accident, so should be safe to delete.

[JakeHillion]

Did these swap for a reason?

[hodgesds]

was thinking going from the least specific to the most specific for ordering, but either way is fine

[JakeHillion]

nit: would it be clearer if we split these into two early returns?

Suggested change

	if (!(cachec = lookup_cache_ctx(cctx->cache_idx)) ||
	!(nodec = lookup_node_ctx(cctx->node_idx)))
	return;
	if (!(cachec = lookup_cache_ctx(cctx->cache_idx)))
	return;
	if (!(nodec = lookup_node_ctx(cctx->node_idx)))
	return;