-
Notifications
You must be signed in to change notification settings - Fork 122
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Support multi completion events: v2 #130
Conversation
This is used to parameterize Op<T, CqeType>. It is now possible to specialize blanket implmentations, such as Future, to classes of Ops
Adds infrastructure to store Completions a slab backed indexed linked list
Provides the infrastructure for recieving multiple CQE's for a single Op's lifecycle
I'm starting to look at your branch. This branch isn't about zero copy commands, its a precursor, I see now. Is there any way of testing this? |
I like the linked list in a slab implementation. I'll note its a doubly linked list and I didn't see the benefit of the backward links but maybe that detail future proofs the list somehow. Is there a branch that shows the CompletionList or the CompletionEntries being passed or consumed by a zero copy operation? I had other comments but the Completion slab is the big addition and I like it. I wonder if @Noah-Kennedy could chime in on it? I could try to change my stream fork to use the same mechanism, getting rid of the Vec of results. |
So both pop and push are 1 lookup. Admittedly, holding the end index is sufficient for this, so the list itself doesn't need to be doubly linked |
Yes, I've rebased the send_zc branch on to this |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Sorry about the delay! I really like this approach. It needs a few finishing touches before it's good to go.
@Noah-Kennedy - I think I've addressed your concerns - let me know if you've more. |
@FrankReh Not at all, I'd noticed we were in different time-zones. I'm also I'm genuinely grateful for the level of scrutiny these Pr's are receiving - these things are generally better got right or close to it first time around. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This is very nice. I think Noah has already said that too. I have a few small issues around comments.
It'd be very easy to add in |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Hope you are fresh when you read these. I think they are all comment related.
@FrankReh All comments addressed |
I'm going to have this page dismiss Noah's review. The page seems broken somehow. When his review is expanded, it shows everything having been resolved. Maybe it just means he was expected to review and accept. I think we can risk it. If more changes are needed, we'll make them with a new PR. |
I believe Noah's issues were handled. I don't believe there is anything else that should block this merge.
Thanks for this really nice work and for your patience and timely turn arounds. I look forward to using these changes myself. |
# 0.4.0 (November 5th, 2022) ### Fixed - Fix panic in Deref/DerefMut for Slice extending into uninitialized part of the buffer ([#52]) - docs: all-features = true ([#84]) - fix fs unit tests to avoid parallelism ([#121]) - Box the socket address to allow moving the Connect future ([#126]) - rt: Fix data race ([#146]) ### Added - Implement fs::File::readv_at()/writev_at() ([#87]) - fs: implement FromRawFd for File ([#89]) - Implement `AsRawFd` for `TcpStream` ([#94]) - net: add TcpListener.local_addr method ([#107]) - net: add TcpStream.write_all ([#111]) - driver: add Builder API as an option to start ([#113]) - Socket and TcpStream shutdown ([#124]) - fs: implement fs::File::from_std ([#131]) - net: implement FromRawFd for TcpStream ([#132]) - fs: implement OpenOptionsExt for OpenOptions ([#133]) - Add NoOp support ([#134]) - Add writev to TcpStream ([#136]) - sync TcpStream, UnixStream and UdpSocket functionality ([#141]) - Add benchmarks for no-op submission ([#144]) - Expose runtime structure ([#148]) ### Changed - driver: batch submit requests and add benchmark ([#78]) - Depend on io-uring version ^0.5.8 ([#153]) ### Internal Improvements - chore: fix clippy lints ([#99]) - io: refactor post-op logic in ops into Completable ([#116]) - Support multi completion events: v2 ([#130]) - simplify driver operation futures ([#139]) - rt: refactor runtime to avoid Rc\<RefCell\<...>> ([#142]) - Remove unused dev-dependencies ([#143]) - chore: types and fields explicitly named ([#149]) - Ignore errors from uring while cleaning up ([#154]) - rt: drop runtime before driver during shutdown ([#155]) - rt: refactor drop logic ([#157]) - rt: fix error when calling block_on twice ([#162]) ### CI changes - chore: update actions/checkout action to v3 ([#90]) - chore: add all-systems-go ci check ([#98]) - chore: add clippy to ci ([#100]) - ci: run cargo test --doc ([#135]) [#52]: #52 [#78]: #78 [#84]: #84 [#87]: #87 [#89]: #89 [#90]: #90 [#94]: #94 [#98]: #98 [#99]: #99 [#100]: #100 [#107]: #107 [#111]: #111 [#113]: #113 [#116]: #116 [#121]: #121 [#124]: #124 [#126]: #126 [#130]: #130 [#131]: #131 [#132]: #132 [#133]: #133 [#134]: #134 [#135]: #135 [#136]: #136 [#139]: #139 [#141]: #141 [#142]: #142 [#143]: #143 [#144]: #144 [#146]: #146 [#148]: #148 [#149]: #149 [#153]: #153 [#154]: #154 [#155]: #155 [#157]: #157 [#162]: #162
@Noah-Kennedy @FrankReh . I took on board your comments with the other PR and wanted to offer an alternative implementation to #128. This addresses some of the shortcomings of #128.
Multi CQE's are now held on a driver-scoped slab of completions. This goes some way to address the biggest shortcoming of the other approach, which was its propensity to allocate if multiple completion events were received for a single Op in quick succession. Obviously, allocation can still occur, but the total storage space is scoped to the runtime rather than per Op - so making it big initially benefits all Ops.
Op is now parameterized over classes of Ops, discriminated with a marker type. This allows us to easily add classes of Operations. All current ones are represented by
SingleCQE
.I shall (haven't yet) rebase Support zero-copy send #123 to show its use.now merged onto this. Will update that with performance figures for my use case.@FrankReh I took on board your comments around Completed + enum variants vs adding another variant.
Lifecycle::CompletionList will never be observed for SingleCQE types.