Skip to content
New issue

Have a question about this project? # for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “#”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? # to your account

[WIP] Use fetch_max instead of bump_atomic_lsn #1233

Closed
wants to merge 1 commit into from
Closed

[WIP] Use fetch_max instead of bump_atomic_lsn #1233

wants to merge 1 commit into from

Conversation

Yam76
Copy link

@Yam76 Yam76 commented Jan 1, 2021

The relevant part of issue rust-lang/rust#48655 was resolved in rust-lang/rust#72324 so this PR should eventually fix the corresponding TODO in pagecache.

This initial PR does some of the grind work of replacing bump_atomic_lsn.

NOTE: I'm not familiar with Rust's atomics, so I don't know if SeqCst is actually what you want.

For cargo test, my OS (Windows 10, WSL) couldn't hole punch the heap, and I got a lot of 60 second timeouts. Also, the tests got stuck on my end at around tree_bug_39, so I had to kill the process. It might be better on your end.

@Yam76 Yam76 changed the title Do the initial work of using fetch_max [WIP] Use fetch_max instead of bump_atomic_lsn Jan 1, 2021
@Yam76 Yam76 marked this pull request as draft January 1, 2021 01:40
@Yam76 Yam76 marked this pull request as ready for review January 1, 2021 01:45
@spacejam
Copy link
Owner

spacejam commented Jan 1, 2021

Hey @Yam76, this is explicitly avoiding fetch_max which did not stabilize until Rust 1.45. sled is currently targeting 1.39 as the minimum supported rust version (MSRV).

@spacejam spacejam closed this Jan 1, 2021
@Yam76
Copy link
Author

Yam76 commented Jan 2, 2021

D'oh, that's what I get for not reading the full README. Thank you for the response.

@spacejam
Copy link
Owner

spacejam commented Jan 2, 2021 via email

# for free to join this conversation on GitHub. Already have an account? # to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants