-
Notifications
You must be signed in to change notification settings - Fork 1.1k
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
Fix the definition of sigevent #2813
Conversation
r? @Amanieu (rust-highfive has picked a reviewer for you, use r? to override) |
@Amanieu what can we do about backwards-compat? It should only be a problem for crates that uses the fairly obscure |
AFAIK no, the cheapest way is |
@Amanieu Semi-related, is there any discussion/progress about libc crate versioning while I wasn't here? I'd like to resolve the root cause of such an issue someday... |
No, there's not been any discussion on this. We could bring it up at the next libs team meeting, but I'm not really sure what needs to be done here. |
Then here's what I suggest:
Does that sound reasonable? Is craterbot already activated for this repo? |
Crater is really designed to test builds of the Rust compiler & standard library, not external crates like libc. I'm sure it would be possible to adapt it, but I'm not familiar with crater and wouldn't know where to start. As an alternative, it's possible to download the source code for all crates published on crates.io: https://twitter.com/m_ou_se/status/1433085053056262144 This would allow grepping for any uses of |
I used a command inspired by that Twitter post to download and search every crate version. What I found was:
|
When Tokio gets event notification from kevent, the pending future may belong to a different thread. If so, it signals the other thread and pends on kevent again. But the other thread may not have had time to call aio_return yet. In that case, the first thread's kevent will immediately return again. The correct solution is to set EV_ONESHOT on the aiocb. But Rust's libc doesn't currently expose the necessary field. Until it does, restrict Tokio to just a single thread. rust-lang/libc#2813
Also, disallow using SigevNotify::SigevThreadId on musl. I don't think it ever worked. Depends on rust-lang/libc#2813 Blocks tokio-rs/tokio#4728
Also, disallow using SigevNotify::SigevThreadId on musl. I don't think it ever worked. Depends on rust-lang/libc#2813 Blocks tokio-rs/tokio#4728
@Amanieu do you approve of the plan to fix libc and deal with the breakage by immediately publishing patch releases of Nix? |
I'm concerned about the amount of breakage: there's a lot of crates that aren't using the latest major version of Nix. Do you think we could implement a backward-compatibility hack using |
We would definitely patch more than just the current major version. I think I should be able to do the top five or six. But what about this DerefMut proposal. How would that work? |
If we implement |
That's a good idea, and it looks like it works! I'll amend the PR. |
@Amanieu the remaining CI failure is due to a limitation in the style checker. It doesn't understand that the type declaration is part of an impl block. I'm not sure how to fix that. Do you have any ideas? |
pub sigev_notify: ::c_int, | ||
pub sigev_signo: ::c_int, | ||
pub sigev_value: ::sigval, | ||
pub sigev_notify_thread_id: ::lwpid_t, |
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.
I would prefer if this was a union
that just contains a single field. This forces the access to the sigev_notify_thread_id
field to be in an unsafe
block, which is necessary since this may be an uninitialized field in the union.
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.
But the whole point of this structure is for backwards-compatibility with older libc versions, and they didn't require unsafe
. Wouldn't it break backwards-compatibility then, if we put this field into a union?
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.
It just needs to be backwards-compatible enough for existing users, which is basically just nix. Does nix access this field outside of an unsafe block?
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.
Yes, existing Nix does not use unsafe
.
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.
Also, I don't think that it should be considered unsafe to access the union member when using the backwards-compatible definition. Since it's defined as a struct with a bunch of padding fields, the whole structure should always be initialized, and it shouldn't be a problem to read from the "union" member.
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.
I think this is most likely fine, but I would like a second opinion from @JohnTitor.
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.
Hmm, I'm not sure if I've got 100% of the context here so don't take my thought as a strong opinion, but I think it doesn't have to be perfectly correct while it's just for back-compat.
Not sure about how to deal with the style checker, @JohnTitor any ideas? |
Fixed in #2824 |
Rebased and squashed. |
This PR has seen a lot of action, and it starts with a note of "...will fail to build with these changes". I skimmed it and am not sure where it's at right now, so I'm marking it as a breakage candidate since we're likely to start evaluating things we want to pull in for a 0.3 release real soon now, and I don't want this to get missed in the review for that, even if it is not breaking anything. However if you could make the PR's opening message match what the current status of this PR is, that would be greatly appreciated. |
Updated, @workingjubilee ! The current status of this PR is that, as far as I am concerned, it is complete. However, there is a trilemma: we cannot have both backwards-compatibility AND rust 1.13.0 support AND an unmodified style checker. The current state of the PR relaxes the style checker, but I would be ok with relaxes either or both of the other conditions, too. |
@workingjubilee as explained in the current PR summary, this change should not cause any breakage. Can we move forward with it? And as far as the trilemma goes, which is your preference to drop?
|
🤔 |
Let me see if I can get a reduced version of #2813 through. Rust 1.13 is my first choice as to what to drop, at the moment. |
@workingjubilee do you need help with said reduced version? |
@workingjubilee I rebased on main. This version retains Rust 1.14.0 compatibility, but relaxes the style checker. |
Also, disallow using SigevNotify::SigevThreadId on musl. I don't think it ever worked. Depends on rust-lang/libc#2813 Blocks tokio-rs/tokio#4728
Also, disallow using SigevNotify::SigevThreadId on musl. I don't think it ever worked. Depends on rust-lang/libc#2813 Blocks tokio-rs/tokio#4728
Also, disallow using SigevNotify::SigevThreadId on musl. I don't think it ever worked. Depends on rust-lang/libc#2813 Blocks tokio-rs/tokio#4728
BTW, I don't think I ever explained the motivation for this change. Without this change, it is impossible to use a multi-threaded Tokio runtime with POSIX AIO. Only a single-threaded runtime can be used. Switching to a multi-threaded runtime by using this libc branch improves my application's performance by 180% . So the PR isn't just for correctness's sake; it's got real value too. |
Also, disallow using SigevNotify::SigevThreadId on musl. I don't think it ever worked. Depends on rust-lang/libc#2813 Blocks tokio-rs/tokio#4728
Also, disallow using SigevNotify::SigevThreadId on musl. I don't think it ever worked. Depends on rust-lang/libc#2813 Blocks tokio-rs/tokio#4728
Also, disallow using SigevNotify::SigevThreadId on musl. I don't think it ever worked. Depends on rust-lang/libc#2813 Blocks tokio-rs/tokio#4728
* Allow setting kevent_flags on struct sigevent Also, disallow using SigevNotify::SigevThreadId on musl. I don't think it ever worked. Depends on rust-lang/libc#2813 Blocks tokio-rs/tokio#4728 * Inline libc::sigevent Because the PR to libc is stalled for over one year, with no sign of progress.
☔ The latest upstream changes (presumably #3487) made this pull request unmergeable. Please resolve the merge conflicts. |
It was originally defined back before rust could represent C unions. So instead of defining the union field correctly, it simply defined that union's most useful field. Define it correctly now. Include a backwards-compatibility mechanism: Rename sigevent's old definition and hide it. Implement Deref and DerefMut from sigevent to the old definition, so consumers will still be able to use the old field name.
sigevent's Debug, PartialEq, and Hash trait impls might read union fields that could be potentially uninitialized by a standard initializer. Those trait impls shouldn't be present (see rust-lang#2816), but can't easily be removed. Until they get removed, the constructor must be `unsafe` to force the user to zero all fields. The same issue applies to the Deref<Target=sigevent_0_2_126> trait impl, which exists only for backwards compatibility.
Rust 1.14.0 and earlier don't correctly expand #[cfg()] attributes within the macro, sometimes necessitating multiple s! macros per file.
If #3630 is a rebase of this PR, we can then close this PR? 🤔 |
Considering the challenges this PR has had and how long it has been open, I think we are better off not worrying about 0.2 and just going forward with a fix for 1.0 via #3630. |
It was originally defined back before rust could represent C unions. So
instead of defining the union field correctly, it simply defined that
union's most useful field. Define it correctly now.
Include a backwards-compatibility mechanism: Rename sigevent's old
definition and hide it. Implement Deref and DerefMut from sigevent to
the old definition, so consumers will still be able to use the old field
name.