Data race? about golang-fifo

[richardartoul]

Your hot fix looks good to me, but I don’t understand the two points you made above. The issue is just that there are two concurrent goroutines loading the same cache value at the same time and they both end up modifying the visited field because they’re not holding an exclusive lock.

another alternative would be to add a sub-lock inside the entry struct and acquire that before marking visited as true. For now though I’d probably merge your existing hot fix since the more complex solutions probably require detailed concurrent benchmarking.

[richardartoul]

I could see this being an intentional performance thing if setting a boolean field compiles down to a single instruction 🤔 but it makes using this library in any codebase that runs tests with the race detector enabled impossible

[scalalang2]

@richardartoul Thank you for reporting.

I've researched one thing that can cause this.

1. Go map is not addressable which means the pointed value may change during performing operations
2. Changes in pointed value can occur when go map rebalances itself to ensure O(1) access.

I think that there are two possible solutions here

• Simply change RLock to Lock to make all operations work sequentially
• Adopt a concurrent-map implementation like this or sync.Map

Hot Fixes

This is a critical bug, I'll handle it with a hotfix and conduct further researchs for this.

[scalalang2]

I also want to hear @richardartoul opinions about this.