-
Notifications
You must be signed in to change notification settings - Fork 2.4k
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
Optimize usage under rustup. #11917
Optimize usage under rustup. #11917
Conversation
cc @rbtcollins |
Failed to set assignee to
|
r? @weihanglo (rustbot has picked a reviewer for you, use r? to override) |
// use hard links to a single binary. If rustup ever changes | ||
// that setup, then I think the worst consequence is that this | ||
// optimization will not work, and it will take the slow path. | ||
if tool_meta.len() != rustup_meta.len() { |
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 seems hazardous. It's entirely possible for two binaries to have the same length without having the same contents. If you're checking for hardlinks, shouldn't this compare ino
and dev
, at least on Unix? (I'm not sure how to improve this on Windows; are we using Windows hardlinks where available?)
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.
Following up: according to @ChrisDenton, you need to open both files, and then while both are open, call GetFileInformationByHandleEx
and make sure both volume and ID are identical.
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.
Essentially you'd need to compare FILE_ID_INFO
structs. Though there are some nuances here and the point about keeping both file handles open is crucial because ids aren't otherwise guaranteed to be stable (see [MS-FSCC] reference). See also this LLVM bug.
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.
That's possible, but I'm concerned about the potential complexity or difficulty in getting that right. There are various situations where symlinks are used, and I can't guarantee that the files won't end up as a copy, or have issues across network mounts, for example. The file sizes are currently an order of magnitude different, and I think the chance for them being the same is very unlikely.
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'm not sure if this matters but the reason the fs footprint of rustup looks quite different in different places is because of android (no hardlink support), bew (everything is a symlink to a symlink from our next release) (and I guess snap and other 3rd-party distributions can also differ to what one might expect from looking at our code).
In particular see rust-lang/rustup#3137 for some context.
tl;dr: there is no guarantee that the proxy and rustup itself are the same file, even though that is our default installation logic.
On android they are symlinks.
On MacOSX with brew they are symlinks from our next release.
And the consequence of the tool not being a proxy is that someone has deliberately placed e.g. a 'rustc' wrapper that does something, which cargo would then not run.
Checking for the same length will detect every common situation where they are different except for two cases I can see: two different binaries, alike in length, and two different symlinks, alike in length.
For binaries I agree - its very unlikely that two different binaries the same length as rustup is large enough that the law of small numbers doesn't really apply.
For symlinks, I suggest doing a readlink on the file. It is cheap enough to still be a lot faster than rustup manifest parsing (which I plan to do something about someday, but its not top of the list, and even after, not running code we don't need to run is how we make things fast).
Oh and the final case - I alluded to above with 'common' - lets exclude other file types from consideration. Special node types should just immediate take the slow path.
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.
For symlinks, I suggest doing a readlink on the file. It is cheap enough to still be a lot faster than rustup manifest parsing (which I plan to do something about someday, but its not top of the list, and even after, not running code we don't need to run is how we make things fast).
Can you say more about why this would be needed? If the proxy symlink points at rustup, shouldn't they have the same size?
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 is possible for the symlink to point at something else with the same length path.
The chance of two unrelated binary lengths colliding when they are ~11M in size (current rustup-init release size on Windows) is pretty low. But the chance of two ~100 byte paths being the same length is much much higher, and then multiply that out by our growing user bases I think its worth mitigating the risk.
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.
Maybe I'm still unclear, but this uses the standard metadata
function which reads the target of a symlink (recursively). If rustup
and rustc
are symlinks to different things (with the same length path), they'll still have different length targets.
If that doesn't resolve your concern, can you show a specific example? For example:
/usr/bin/rustc -> /usr/bin/rust-compiler
/usr/bin/rustup -> /usr/bin/rustup-thingy
/usr/bin/rust-compiler 669176
/usr/bin/rustup-thingy 8027337
These have the same length paths, but different length targets, so they should be treated as being different.
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.
ok, so you're using fs::metadata(path).len(), not symlink_metadata ? Then I'm fine with that as-is
// This is an optimization to circumvent the rustup proxies | ||
// which can have a significant performance hit. The goal here | ||
// is to determine if calling `rustc` from PATH would end up | ||
// calling the proxies. | ||
// | ||
// This is somewhat cautious trying to determine if it is safe | ||
// to circumvent rustup, because there are some situations | ||
// where users may do things like modify PATH, call cargo | ||
// directly, use a custom rustup toolchain link without a | ||
// cargo executable, etc. However, there is still some risk | ||
// this may make the wrong decision in unusual circumstances. | ||
// | ||
// First, we must be running under rustup in the first place. |
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.
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.
Querying rustc is certainly a possibility, but the approach taken there incurs additional startup time for an initial cache. Adding a new flag to rustc has a fairly high bar, but adding a transparent optimization in cargo has no user-facing interaction so should be easier to move forward with. So, in terms of the global complexity (changes to cargo, rustc, and/or rustup, user-facing documentation, etc.), this seemed like the simplest solution with the least risk, and can receive benefits immediately rather than waiting a potentially very long time.
If this solution ends up having issues that one of the other solutions could address, then I think it would be worthwhile to re-investigate a different approach.
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.
Adding a new flag to rustc has a fairly high bar
Even if its just a new enumerated value for an existing flag?
If this solution ends up having issues that one of the other solutions could address, then I think it would be worthwhile to re-investigate a different approach.
If I understand correctly, this solution could start failing and we'd never know it, right?
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.
Even if its just a new enumerated value for an existing flag?
Yea, new options are almost always added as an unstable option, and then it needs to go through the process of making a case for the compiler team to stabilize.
If I understand correctly, this solution could start failing and we'd never know it, right?
It is possible, though I think unlikely in most cases. I think any major regressions would require a significant change in the design of rustup, and I think that is unlikely for the foreseeable future. I could add a test that requires rustup to be installed if that may help with that concern.
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 can add in that rustup has been consulted and we quite like this approach.
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 definitely prefer a solution like this that doesn't add any additional invocations of rustc.
839d8d6
to
7263f3b
Compare
// First, we must be running under rustup in the first place. | ||
let toolchain = self.get_env_os("RUSTUP_TOOLCHAIN")?; | ||
// If the tool on PATH is the same as `rustup` on path, then | ||
// there is pretty good evidence that it will be a proxy. |
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.
We have an exact list of the proxies we offer btw. I think it is a good idea to only take the fastpath for things we are known to proxy. I'm happy to commit to keeping a copy of that list in Cargo up to date. New proxies are very rare.
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.
Cargo is currently hard-coded to only use this for rustc
and rustdoc
. I added an assert to validate that requirement, and I think we can extend it in the future if needed. I don't think we quite yet need to have an exhaustive list for all the proxies (just to keep things simple for now).
I'm in favour of this as discussed. One note - I'm looking to sharply constrain toolchain path based toolchains in the next release - removing the relative-path support entirely (or at least feature-flagging it to drive some feedback), and in a later release I hope to remove it. There's no tracking bug yet but some discord discussion. All of which to say, I don't think you should worry about path based toolchains, other than rejecting toolchains that contain '/' or ''. |
let toolchain_exe = home::rustup_home() | ||
.ok()? | ||
.join("toolchains") | ||
.join(&toolchain) |
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 could look at pretty random places on the filesystem with a path based toolchain. I suggest breaking out of this logic early based on toolchain == 'none' || toolchain.contains('/') || toolchain.contains('\\')
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.
Good catch! I have added that.
Another complication but probably ok: Users can specify I think this is ok because if someone runs cargo without the proxy, with that variable set, the directory won't exist, and the fallback path of just invoking the rustc proxy will invoke a rustup proxy, which will then perform resolution, ending up with the explicitly requested toolchain being run. If the user is running a cargo from a different toolchain, directly, with a RUSTUP_TOOLCHAIN variable set, then I think they can keep both pieces. |
@weihanglo Just checking in to see if you have any questions or thoughts on this. I realize this PR contains a lot of words for a relatively small code change. |
src/cargo/util/config/mod.rs
Outdated
// assert is to ensure that if it is ever used for something else in | ||
// the future that you must ensure that it is a proxy-able tool, or if | ||
// not then you need to use `maybe_get_tool` instead. | ||
assert!(matches!(tool, "rustc" | "rustdoc")); |
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.
(nit) Should change this to an enum instead of runtime assertion, and also update the function doc comment?
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.
Sure, I went ahead and added an enum.
src/cargo/util/config/mod.rs
Outdated
if toolchain_exe.exists() { | ||
Some(toolchain_exe) | ||
} else { | ||
None | ||
} |
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.
(nit)
if toolchain_exe.exists() { | |
Some(toolchain_exe) | |
} else { | |
None | |
} | |
toolchain_exe.exists().then_some(toolchain_exe) |
} | ||
// Try to find the tool in rustup's toolchain directory. | ||
let tool_exe = Path::new(tool).with_extension(env::consts::EXE_EXTENSION); | ||
let toolchain_exe = home::rustup_home() |
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.
home
crate still access RUSTUP_HOME
via std::env
. Will there be an inconsistency when people set their RUSTUP_HOME
in [env]
?
Line 93 in 39684ff
match env.var_os("RUSTUP_HOME").filter(|h| !h.is_empty()) { |
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 shouldn't under normal circumstances, since the rustup proxies set RUSTUP_HOME
, the [env]
value will be ignored.
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 it would be a good idea to prohibit this before stablisation @ehuss
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.
Sure, I went ahead and posted #12101.
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 looks pretty good and simpler than other alternatives to me. Given that rustup folks agree on this change, Feel free to r=weihanglo
if there is no further discussion needed with them.
I think this should be ready to go. My intent is to get this out on nightly and hopefully get enough real-world testing to determine if there is a problem, and we can either fix it or back it out. Unfortunately nightly doesn't get enough testing in some of the more unusual environments, but hopefully it will be enough to be moderately confident it should be ok. @bors r=weihanglo |
☀️ Test successful - checks-actions |
Update cargo 10 commits in ac84010322a31f4a581dafe26258aa4ac8dea9cd..569b648b5831ae8a515e90c80843a5287c3304ef 2023-05-02 13:41:16 +0000 to 2023-05-05 15:49:44 +0000 - xtask-unpublished: output a markdown table (rust-lang/cargo#12085) - fix: hack around `libsysroot` instead of `libtest` (rust-lang/cargo#12088) - Optimize usage under rustup. (rust-lang/cargo#11917) - Update lock to normalize `home` dep (rust-lang/cargo#12084) - fix: doc-test failures (rust-lang/cargo#12055) - feat(cargo-metadata): add `workspace_default_members` (rust-lang/cargo#11978) - doc: clarify implications of `cargo-yank` (rust-lang/cargo#11862) - chore: Use `[workspace.dependencies]` (rust-lang/cargo#12057) - support for shallow clones and fetches with `gitoxide` (rust-lang/cargo#11840) - Build by PackageIdSpec, not name, to avoid ambiguity (rust-lang/cargo#12015) r? `@ghost`
Closes #10986
This optimizes cargo when running under rustup to circumvent the rustup proxies. The rustup proxies introduce overhead that can make a noticeable difference.
The solution here is to identify if cargo would normally run
rustc
from PATH, and the currentrustc
in PATH points to something that looks like a rustup proxy (by comparing it to therustup
binary which is a hard-link to the proxy). If it detects this situation, then it looks for a binary in$RUSTUP_HOME/toolchains/$TOOLCHAIN/bin/$TOOL
. If it finds the direct toolchain executable, then it uses that instead.Considerations
There have been some past attempts in the past to address this, but it has been a tricky problem to solve. This change has some risk because cargo is attempting to guess what the user and rustup wants, and it may guess wrong. Here are some considerations and risks for this:
Setting
RUSTC
(as in Set RUSTC and RUSTDOC env for child processes run through the proxy rustup#2958) isn't an option. This makes theRUSTC
setting "sticky" through invocations of different toolchains, such as a cargo subcommand or build script which does something likecargo +nightly build
.Changing
PATH
isn't an option, due to issues like rustup 1.25: On Windows, nested cargo invocation with a toolchain specified fails rustup#3036 where cargo subcommands would be unable to execute proxies (so things like+toolchain
shorthands don't work).Setting other environment variables in rustup (as in Add RUSTUP_TOOLCHAIN_DIR rustup#3207 which adds
RUSTUP_TOOLCHAIN_DIR
the path to the toolchain dir) comes with various complications, as there is risk that the environment variables could get out of sync with one another (like withRUSTUP_TOOLCHAIN
), causing tools to break or become confused.There was some consideration in that PR for adding protections by using an encoded environment variable that could be cross-checked, but I have concerns about the complexity of the solution.
We may want to go with this solution in the long run, but I would like to try a short term solution in this PR first to see how it turns out.
This won't work for a
rustup-toolchain.toml
override with apath
setting. Cargo will use the slow path in that case. In theory it could try to detect this situation, which may be an exercise for the future.Some build-scripts, proc-macros, or custom cargo subcommands may be doing unusual things that interfere with the assumptions made in this PR. For example, a custom subcommand could call a
cargo
executable that is not managed by rustup. Proc-macros may be executing cargo or rustc, assuming it will reach some particular toolchain. It can be difficult to predict what unusual ways cargo and rustc are being used. This PR (and its tests) tries to make extra sure that it is resilient even in unusual circumstances.The "dev" fallback in rustup can introduce some complications for some solutions to this problem. If a rustup toolchain does not have cargo, such as with a developer "toolchain link", then rustup will automatically call either the nightly, beta, or stable cargo if they are available. This PR should work correctly, since rustup sets the correct
RUSTUP_TOOLCHAIN
environment variable for the actual toolchain, not the one where cargo was executed from.Special care should be considered for dynamic linking.
LD_LIBRARY_PATH
(linux),DYLD_LIBRARY_PATH
(macos), andPATH
(windows) need to be carefully set so thatrustc
can find its shared libraries. Directly executingrustc
has some risk that it will load the wrong shared libraries. There are some mitigations for this. macOS and Linux use rpath, and Windows looks in the same directory asrustc.exe
. Also, rustup configures the dyld environment variables from the outer cargo. Finally, cargo also configures these (particularly for the deprecated compiler plugins).This shouldn't impact installations that don't use rustup.
I've done a variety of testing on the big three platforms, but certainly nowhere exhaustive.
There is risk about future rustup versions changing some assumptions made here. Some assumptions:
RUSTUP_TOOLCHAIN
is set, then the proxy always runs exactly that toolchain and no other. If this changes, cargo could execute the wrong version. CurrentlyRUSTUP_TOOLCHAIN
is the highest priority toolchain override and is fundamental to how toolchain selection becomes "sticky", so I think it is unlikely to change.RUSTUP_TOOLCHAIN
to a value that is exactly equal to the name of the toolchain in thetoolchains
directory. This works for user shorthands likeRUSTUP_TOOLCHAIN=nightly
, which gets converted to the full toolchain name. However, it does not work forpath
overrides (see above).toolchains
directory layout is always$RUSTUP_HOME/toolchains/$TOOLCHAIN
. If this changes, then I think the only consequence is that cargo will go back to the slow path.My hope is that if assumptions are no longer valid that the worst case is that cargo falls back to the slow path of running the proxy from PATH.
Performance
This change won't affect the performance on Windows because rustup currently alters PATH to point to the toolchain directory. However, rust-lang/rustup#3178 is attempting to remove that, so this PR will be required to avoid a performance penalty on Windows. That change is currently opt-in, and will likely take a long while to roll out since it won't be released until after the next release, and may be difficult to get sufficient testing.
I have done some rough performance testing on macOS, Windows, and Linux on a variety of different kinds of projects with different commands. The following attempts to summarize what I saw.
The timings are going to be heavily dependent on the system and the project. These are the values I get on my systems, but will likely be very different for everyone else.
The Windows tests were performed with a custom build of rustup with rust-lang/rustup#3178 applied and enabled (stock rustup shows no change in performance as explained above).
The data is summarized in this spreadsheet: https://docs.google.com/spreadsheets/d/1zSvU1fQ0uSELxv3VqWmegGBhbLR-8_KUkyIzCIk21X0/edit?usp=sharing
hello-world
has a particularly large impact of about 1.68 to 2.7x faster. However, a large portion of this overhead is related to runningrustc
at the start to discover its version and querying it for information. This is cached after the first run, so except for first-time builds, the effect isn't as noticeable. The "check with info" row is an benchmark that removestarget/debug/deps
but keeps the.rustc_info.json
file.Incremental builds are a bit more difficult to construct since it requires customizing the commands for each project. I only did an incremental test for cargo itself, running
touch src/cargo/lib.rs
and thencargo check --lib
.These measurements excluded the initial overhead of launching the rustup proxy to launch the initial cargo process. This was done just for simplicity, but it makes the test a little less characteristic of a typical usage, which will have some constant overhead for running the proxy.
These tests were done using
hyperfine
version 1.16.1. The macOS system was an M2 Max (12-thread). The Windows and Linux experiments were run on a AMD Ryzen Threadripper 2950X (32-thread). Rust 1.68.2 was used for testing. I can share the commands if people want to see them.