Internalize uring #285
Internalize uring #285
Conversation
This prepares us for the release of a new version of liburing. iou and uring-sys changed a bit, in particular iou. Most of the changes stem from the fact that we now can (should) use the nix versions of things like MsgFlags. The registrar and sqe apis are also marginally different and are updated Note that this patch, on itself, will not compile as it needs matching iou and uring-sys. The next patch will fix that as it brings those dependencies inside glommio.
glommer
force-pushed
the
internalize-uring
branch
from
c5dede7
to
a115e8b
Compare
| unsigned int nbytes, | ||
| unsigned int splice_flags) | ||
| { | ||
| io_uring_prep_splice(sqe, fd_in, fd_out, off_in, off_out, nbytes, splice_flags); |
There was a problem hiding this comment.
There's a bug with the argument order (fd_in, off_in, not fd_in, fd_out) (ringbahn/uring-sys#25), maybe it's worth addressing it here right away.
There was a problem hiding this comment.
Thanks for the review!
I would prefer not to touch this for now because we don't use splice. That said, if uring-sys keeps being unmaintained then maybe we will. Let's leave that on hold for a while.
| ) -> Option<SQEs<'a>> { | ||
| atomic::fence(Ordering::Acquire); | ||
|
|
||
| let head: u32 = *sq.khead; |
There was a problem hiding this comment.
There's a possible issue with prepare_sqes and thread safety here that should be addressed: ringbahn/iou#61
There was a problem hiding this comment.
The issue indicates that the issue only happens with sqpool mode, right?
We don't use that, so I'll take the same approach
|
Thanks a lot! The .swp file was on purpose, just so I could tell the world how I love vim. I prefer usizes for higher layer returns (like wait) ,but iou, which used usizes, change most of it (but not all, I have a consistency issue opened with them for the buffer ids) to u32. My idea is to translate from u32 to usize before returning from |
|
That makes sense, |
|
EDIT: just saw that this was commented on already.. |
Those crates are moving in a fairly slow pace, while uring moves fast. As much as I fully understand that the maintainers have other commitments, this is harming us a bit. Importing this inside glommio is, actually a way to avoid a fork in my view: I consider this temporary and would encourage anyone that is sending code that touches iou and uring-sys to make a bona fide effort to upstream them. But by having them as a buffer here, we can detach our lifetimes.
glommer
force-pushed
the
internalize-uring
branch
from
93ecc66
to
447a38e
Compare
|
Thank you both for the review |
What does this PR do?
Internalizes uring-sys and iou, hopefully in a temporary manner, until all the code we need makes into their crates release
Motivation
Those crates are moving too slowly, and blocking our progress in a significant manner
Related issues
#283
#243