minhbq-99 / store-buffering-experiment

Observe non-SC behavior in TSO

Store buffering experiments

While learning about memory consistency model, I want to observe the non sequential consistency (SC) behavior on x86 total store ordering (TSO). Normally, we can write litmus test and use herd tool suite herdtools7 to do so. Here is the store buffering litmus test

X86 SB
"Fre PodWR Fre PodWR"
{ x=0; y=0; }
 P0          | P1          ;
 MOV [x],$1  | MOV [y],$1  ;
 MOV EAX,[y] | MOV EAX,[x] ;
locations [x;y;]
exists (0:EAX=0 /\ 1:EAX=0)

However, I want to play around and observe the non-SC behavior using my own code :)

I mimic the litmus test in 2 ways:

• Using 2 concurrent threads
• Using 2 KVM vCPUs

1. Using threads

• Actually this is how litmus test is run using herd tool suite. My implementation is inspired a lot from the generated C code of litmus test by herd tool suite
• The herd tool suite generated test is more robust and run faster than mine as I don't try to run tests concurrently
• Run 500000 tests with this approach I observe the non-SC behavior 9 times on my machine (the number varies a lot, my main goal is to observe non-SC behavior so I don't try to get the average number here)

2. Using KVM vCPUs

• Curious to run the tests on bare CPU without much OS intervention, I play with KVM API to create 2 vCPUs running in real mode
• This approach run tests much slower than the previous one maybe because I do too much switch between host and guest, but it requires fewer tests to observe non-SC behavior
• Run 5000 tests I observe the non-SC behavior 10 times

Update:

• I use perf to find out much of the running time is used to create new threads. In this approach, I use 2 threads only to monitor the 2 vCPUs so I think I don't need to re-create threads per test running, just reset the vCPUs after each test. With the optimization, this approach runs super fast, x15 faster than the first approach and observe more non-SC behaviors. But the non-SC behaviors are so much more than the unoptimized one which makes me worry about the correctness of my implementation. However, I sanity check with fence between the store and load in litmus test, then I cannot observe non-SC behavior as expected. If you find any mistakes in the implemation, please let me know
• Run 500000 tests with the optimized one I observe non-SC behavior 40000 times (not the average number)

Happy learning!