Implement waitable_atomic for FreeBSD #1607
Conversation
Based on the Apple implementation rewritten to use _umtx_op syscall.
|
Thanks for making this pull request! We just need a couple of things from you to start our consideration of the PR. Can you please:
|
Thanks for your quick response. |
|
Sorry for the delay in writing back. My team owns this code and we have been debating internally what to do with this PR. On the one hand, we appreciate the effort to upstream support for other platforms. But on the other, we are hesitant to accept maintenance of code that isn't executed in our CI system and that we have limited ability to test. Our preference would probably be for you to maintain this as an out of tree patch. How much of a hardship would that be for you? As for the code, I wanted to provide some feedback to ensure that you have the best version, regardless of whether it gets merged or maintained out-of-tree. Based on my reading of https://man.freebsd.org/cgi/man.cgi?query=_umtx_op&sektion=2&n=1, I think it would be better to base this off of the linux version rather than the apple version of the code. It seems like |
|
Thanks for the answer. I understand the suggestion for the implementation with absolute time. I will take a look at that in the near future. Maintaining the patches out of tree is doable. As I'm doing it for quite some time now. I just thought it would be nice to upstream some of this code. As it is a working port of MongoDB on a popular posix OS. Of course your team is in the lead and I'm happy with what you decide is possible. |
Similar to the Linux implementation instead of the MacOS version.
Reimplement src/mongo/platform/waitable_atomic.cpp using absolute time. This uses more simple conversions. The implementation is now similar to the Linux version instead of the MacOS version. It should not have any functional change. Reported by: https://infosec.exchange/@david_chisnall/113797905538197081 Reported by: mongodb/mongo#1607 (comment)
|
I'm going to go ahead and close this. I'd rather we not merge non-trivial code that we don't have any ability to test in an ongoing manner in CI. |
Includes patch to implement waitable atomics for FreeBSD for https://jira.mongodb.org/browse/SERVER-81797. This patch is developed by me. It did not go through QA of MongoDB Inc. I asked to upstream the patch but they don't want to take responsibilty for code they don't run themselves. (Which I can understand from their perspective.) mongodb/mongo#1607 (comment) https://jira.mongodb.org/browse/SERVER-99225 Anyway, this has run without errors for a while on my Raspberry Pi + net-mgmt/unifi9. I don't have full QA in place and could only do limited testing. I can't take any responsibilty for the quality of the patch. Make backups and test for yourself in your setup before upgrading any critical production environment. Changes: https://www.mongodb.com/docs/manual/release-notes/7.0/#7.0.23---aug-13--2025 Announcement: https://www.mongodb.com/community/forums/t/mongodb-7-0-23-is-released/327242?u=r_k
Be pedantic that '!= 0' is not the same as '== -1'. And indeed -1 is what is documented in the man page of _umtx_op to indicate an error. No functional change expected. Reported by: mongodb/mongo#1607 (review)
3 participants
Based on the Apple implementation rewritten to use _umtx_op syscall.