Remove barriers and spinlock in epoch_enter and epoch_exit

[Alan-Jowett]

Description

Switch epoch logic from relying on spin-lock for protecting to relying on running at IRQL == DISPATCH_LEVEL.

Batching == 1

Before:

C:\bpf_perf>RelWithDebInfo\bpf_performance_runner.exe -i tests.yml -e .sys -t "Baseline" -b 1 -c  50000000
Test,CPU 0,CPU 1,CPU 2,CPU 3,CPU 4,CPU 5
Baseline,119,119,115,117,116,118

After:

C:\bpf_perf>RelWithDebInfo\bpf_performance_runner.exe -i tests.yml -e .sys -t "Baseline" -b 1 -c  50000000
Test,CPU 0,CPU 1,CPU 2,CPU 3,CPU 4,CPU 5
Baseline,31,30,31,31,31,33

Batching == 64

Before:

C:\bpf_perf>RelWithDebInfo\bpf_performance_runner.exe -i tests.yml -e .sys -t "Baseline" -c  50000000
Test,CPU 0,CPU 1,CPU 2,CPU 3,CPU 4,CPU 5
Baseline,10,11,10,9,10,10

After:

C:\bpf_perf>RelWithDebInfo\bpf_performance_runner.exe -i tests.yml -e .sys -t "Baseline" -c  50000000
Test,CPU 0,CPU 1,CPU 2,CPU 3,CPU 4,CPU 5
Baseline,8,7,7,8,8,8

Testing

CI/CD + performance + stress.

Documentation

Yes.

Installation

No.

Resolves: #2787

[Alan-Jowett]

After generating SPGO data and applying it:

Old epoch:

C:\test>RelWithDebInfo\bpf_performance_runner.exe -i tests.yml -e .sys -t "Baseline" -b 1
Test,CPU 0,CPU 1,CPU 2,CPU 3,CPU 4,CPU 5
Baseline,107,109,108,109,111,113

C:\test>RelWithDebInfo\bpf_performance_runner.exe -i tests.yml -e .sys -t "Baseline"
Test,CPU 0,CPU 1,CPU 2,CPU 3,CPU 4,CPU 5
Baseline,7,6,6,7,6,6

New epoch:

C:\test>RelWithDebInfo\bpf_performance_runner.exe -i tests.yml -e .sys -t "Baseline" -b 1
Test,CPU 0,CPU 1,CPU 2,CPU 3,CPU 4,CPU 5
Baseline,31,28,29,30,30,25

C:\test>RelWithDebInfo\bpf_performance_runner.exe -i tests.yml -e .sys -t "Baseline"
Test,CPU 0,CPU 1,CPU 2,CPU 3,CPU 4,CPU 5
Baseline,4,5,5,5,5,5

[dthaler]

Ok but see suggestions.

[Alan-Jowett]

Blocked on: #2783 as it has the updated version of usersim with kevent.

[codecov]

Codecov Report

Merging #2796 (bddd81e) into main (ebc4816) will decrease coverage by 0.22%.
Report is 2 commits behind head on main.
The diff coverage is 95.39%.

@@            Coverage Diff             @@
##             main    #2796      +/-   ##
==========================================
- Coverage   86.71%   86.50%   -0.22%     
==========================================
  Files         140      140              
  Lines       25894    25926      +32     
==========================================
- Hits        22454    22426      -28     
- Misses       3440     3500      +60     

Files Changed	Coverage Δ
libs/execution_context/ebpf_core.c	91.51% <75.00%> (-0.07%)	⬇️
libs/execution_context/ebpf_program.c	87.28% <75.00%> (+0.21%)	⬆️
libs/runtime/ebpf_epoch.c	95.94% <97.22%> (+<0.01%)	⬆️
libs/execution_context/ebpf_link.c	91.14% <100.00%> (ø)
libs/runtime/unit/platform_unit_test.cpp	99.69% <100.00%> (+<0.01%)	⬆️

... and 6 files with indirect coverage changes

[shankarseal]

My concern with this changes is that this may frequently queue DPCs to cores thereby stalling other DPCs generated by interrupts such as arriving network packets. It will be necessary to test these changes using XDP tests with high bandwidth of inbound traffic and also with other multithreaded stress scenarios to ensure there are no regressions.

an alternative was discussed here and eventully rejected. #2844