Fix building under UML #881

sulix · 2022-09-12T12:22:45Z

At some point, UML (arch/um) support got broken.

We were able to work out the two causes of this at Kangrejos:

A strange SSE2-related SIGTRAP in rustc.
The relocation model needs to be PIE for some reason on UML builds. Maybe this is something better fixed in UML, but it was easier to change here. This is basically a revert of this commit in PR Target specification cleanup #706 — despite the comment there, I couldn't override the command-line version in a per-architecture way.

There's almost definitely a better way of fixing these, and even these fixes would need to be better understood and documented, but I'm putting this out there so we have some documentation of the problems, and something we can apply to test against.

scripts/generate_rust_target.rs

Explain how to run unit tests and documentation tests. Note that the documentation uses "--arch=x86_64" to run KUnit tests because UML is not working at the moment [1]. [1] Rust-for-Linux#881 Signed-off-by: José Expósito <[email protected]>

Explain how to run unit tests and documentation tests. Note that the documentation uses "--arch=x86_64" to run KUnit tests because UML is not working at the moment [1]. [1] Rust-for-Linux/linux#881 Signed-off-by: José Expósito <[email protected]>

Explain how to run unit tests and documentation tests. Note that the documentation uses "--arch=x86_64" to run KUnit tests because UML is not working at the moment [1]. [1] Rust-for-Linux/linux#881 Reviewed-by: David Gow <[email protected]> Signed-off-by: José Expósito <[email protected]>

Explain how to run unit tests and documentation tests. Note that the documentation uses "--arch=x86_64" to run KUnit tests because UML is not working at the moment [1]. [1] Rust-for-Linux#881 Reviewed-by: David Gow <[email protected]> Signed-off-by: José Expósito <[email protected]>

sulix · 2022-12-08T07:20:41Z

As alluded to in ClangBuiltLinux#1715 -- UML is very particular about its linker options, so I suspect it'd take a fair bit of work (and more knowledge than I have) to fix this on the UML side.

I've rebased it and verified it still works, so if there's no objection to working around this on the Rust side, and things like disabling SSE/SSE2 don't break anything, I think it makes sense to merge this for now.

That'll also remove the need to recommend --arch x86_64 in #935.

bjorn3

LGTM

Add a .kunitconfig file with the required configuration options to allow to easily run the KUnit tests without adding them manually: $ ./tools/testing/kunit/kunit.py run --kunitconfig=rust \ --make_options LLVM=1 --arch=x86_64 Note that "CONFIG_UML" is set to "n" because UML is not working at the moment [1]. [1] Rust-for-Linux/linux#881 Reviewed-by: David Gow <[email protected]> Signed-off-by: José Expósito <[email protected]>

Explain how to run unit tests and documentation tests. Note that the documentation uses "--arch=x86_64" to run KUnit tests because UML is not working at the moment [1]. [1] Rust-for-Linux/linux#881 Reviewed-by: David Gow <[email protected]> Signed-off-by: José Expósito <[email protected]>

Add a .kunitconfig file with the required configuration options to allow to easily run the KUnit tests without adding them manually: $ ./tools/testing/kunit/kunit.py run --kunitconfig=rust \ --make_options LLVM=1 --arch=x86_64 Note that "CONFIG_UML" is set to "n" because UML is not working at the moment [1]. [1] Rust-for-Linux#881 Reviewed-by: David Gow <[email protected]> Signed-off-by: José Expósito <[email protected]>

Explain how to run unit tests and documentation tests. Note that the documentation uses "--arch=x86_64" to run KUnit tests because UML is not working at the moment [1]. [1] Rust-for-Linux#881 Reviewed-by: David Gow <[email protected]> Signed-off-by: José Expósito <[email protected]>

ojeda · 2022-12-08T20:12:05Z

scripts/generate_rust_target.rs

+    if cfg.has("UML") {
+        ts.push("relocation-model", "pie");
+    } else {
+        ts.push("relocation-model", "static");
+    }


Since using target spec files is considered unstable inrustc, we are trying to use as many (stable) flags as possible instead, in hopes that, eventually, we will be have an easier time to move to stable-only builds. Could you please override the relocation model via arch/um/Makefile with KBUILD_RUSTFLAGS += ?

ojeda · 2022-12-09T21:52:08Z

scripts/generate_rust_target.rs

@@ -204,7 +204,7 @@ fn main() {
            "data-layout",
            "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128",
        );
-        let mut features = "-3dnow,-3dnowa,-mmx,+soft-float".to_string();
+        let mut features = "-3dnow,-3dnowa,-mmx,+soft-float,-sse,-sse2".to_string();


SSE2 is already disabled already through the arch/x86/Makefile:

KBUILD_RUSTFLAGS += -Ctarget-feature=-sse,-sse2,-sse3,-ssse3,-sse4.1,-sse4.2,-avx,-avx2

But of course that doesn't get picked up by ARCH=um... This seems to work:

KBUILD_RUSTFLAGS += -Ctarget-feature=-sse

in arch/um/Makefile (-sse seems enough), probably with a subarch conditional. That way, it is more localized (since it is a change for UML) and we avoid the unstable target spec file. Or, if UML is supposed to use the same flags, we can just copy the entire line.

I minimized our case down to https://godbolt.org/z/vWcadW96h, and noticed it doesn't happen anymore starting with rustc 1.65.0. Since that version upgraded to LLVM 15 in rust-lang/rust#99464, I guess we can assume it got fixed on LLVM's side (or somehow it doesn't trigger it anymore).

And indeed, with rustc 1.65.0 I could get pass that part compiling the kernel.

But again, we probably want the same flags as x86 anyway, right?

The kernel disables all SSE and similar FP/SIMD instructions on x86-based architectures (partly because we shouldn't be using floats in the kernel, and partly to avoid the need for stack alignment, see: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53383 ) UML does not do the same thing, which isn't in itself a problem, but does add to the list of differences between UML and "normal" x86 builds. In addition, there was a crash bug with LLVM < 14 / rustc < 1.65 when building with SSE, so disabling it fixes rust builds with earlier compiler versions, see: Rust-for-Linux#881 Signed-off-by: David Gow <[email protected]>

sulix · 2022-12-15T15:03:23Z

Thanks, Miguel.

I've updated the series to move everything to the various Makefiles.

Let me know if you'd rather I send this (or something similar) in via the UML mailing list / tree.

ojeda · 2022-12-15T18:02:46Z

Thanks a lot! This will make it simpler later on for us. Looks good to me -- a couple things I noticed:

I think the second commit message should say "LLVM < 15" instead of "LLVM < 14", i.e. LLVM 15 appears to not trigger the issue anymore.
Nit: elsewhere (e.g. the main kernel Makefile) I have been putting RUSTFLAGS closer to CFLAGS (as "language compilers"), i.e. before AFLAGS/LDFLAGS.

Let me know if you'd rather I send this (or something similar) in via the UML mailing list / tree.

It would be great if the UML maintainers take it, actually, so that they start to get involved etc. But if they don't want, I can also route it through the Rust tree. Feel free to email both mailing lists!

By the way, it would be also a good opportunity to see if they want UML listed in Documentation/rust/arch-support.rst etc.

UML expects a position independent executable for some reason, so tell rustc to generate pie objects. Otherwise we get a bunch of relocations we can't deal with in libcore. Signed-off-by: David Gow <[email protected]>

The kernel disables all SSE and similar FP/SIMD instructions on x86-based architectures (partly because we shouldn't be using floats in the kernel, and partly to avoid the need for stack alignment, see: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53383 ) UML does not do the same thing, which isn't in itself a problem, but does add to the list of differences between UML and "normal" x86 builds. In addition, there was a crash bug with LLVM < 15 / rustc < 1.65 when building with SSE, so disabling it fixes rust builds with earlier compiler versions, see: Rust-for-Linux#881 Signed-off-by: David Gow <[email protected]>

sulix · 2022-12-17T04:52:40Z

I've sent this out (along with the CONFIG_HAVE_RUST patch from #766) to the UML list:
https://lore.kernel.org/rust-for-linux/[email protected]/T/

The kernel disables all SSE and similar FP/SIMD instructions on x86-based architectures (partly because we shouldn't be using floats in the kernel, and partly to avoid the need for stack alignment, see: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53383 ) UML does not do the same thing, which isn't in itself a problem, but does add to the list of differences between UML and "normal" x86 builds. In addition, there was a crash bug with LLVM < 15 / rustc < 1.65 when building with SSE, so disabling it fixes rust builds with earlier compiler versions, see: Rust-for-Linux#881 Signed-off-by: David Gow <[email protected]> Reviewed-by: Sergio González Collado <[email protected]> Signed-off-by: Richard Weinberger <[email protected]>

sulix · 2023-02-13T03:16:22Z

These patches are currently sitting in the UML next tree, and so probably will be upstream in 6.3:

ojeda · 2023-02-13T10:40:46Z

Yeah, let's merge this to make rust closer then (we recently decided we will keep the rust branch around until the diff goes to zero), since the one from #766 is in too.

[ Upstream commit 8849818 ] The kernel disables all SSE and similar FP/SIMD instructions on x86-based architectures (partly because we shouldn't be using floats in the kernel, and partly to avoid the need for stack alignment, see: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53383 ) UML does not do the same thing, which isn't in itself a problem, but does add to the list of differences between UML and "normal" x86 builds. In addition, there was a crash bug with LLVM < 15 / rustc < 1.65 when building with SSE, so disabling it fixes rust builds with earlier compiler versions, see: Rust-for-Linux/linux#881 Signed-off-by: David Gow <[email protected]> Reviewed-by: Sergio González Collado <[email protected]> Signed-off-by: Richard Weinberger <[email protected]> Signed-off-by: Sasha Levin <[email protected]>

[ Upstream commit 8849818679478933dd1d9718741f4daa3f4e8b86 ] The kernel disables all SSE and similar FP/SIMD instructions on x86-based architectures (partly because we shouldn't be using floats in the kernel, and partly to avoid the need for stack alignment, see: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53383 ) UML does not do the same thing, which isn't in itself a problem, but does add to the list of differences between UML and "normal" x86 builds. In addition, there was a crash bug with LLVM < 15 / rustc < 1.65 when building with SSE, so disabling it fixes rust builds with earlier compiler versions, see: Rust-for-Linux/linux#881 Signed-off-by: David Gow <[email protected]> Reviewed-by: Sergio González Collado <[email protected]> Signed-off-by: Richard Weinberger <[email protected]> Signed-off-by: Sasha Levin <[email protected]>

[ Upstream commit 8849818 ] The kernel disables all SSE and similar FP/SIMD instructions on x86-based architectures (partly because we shouldn't be using floats in the kernel, and partly to avoid the need for stack alignment, see: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53383 ) UML does not do the same thing, which isn't in itself a problem, but does add to the list of differences between UML and "normal" x86 builds. In addition, there was a crash bug with LLVM < 15 / rustc < 1.65 when building with SSE, so disabling it fixes rust builds with earlier compiler versions, see: Rust-for-Linux/linux#881 Signed-off-by: David Gow <[email protected]> Reviewed-by: Sergio González Collado <[email protected]> Signed-off-by: Richard Weinberger <[email protected]> Signed-off-by: Sasha Levin <[email protected]>

[ Upstream commit 8849818679478933dd1d9718741f4daa3f4e8b86 ] The kernel disables all SSE and similar FP/SIMD instructions on x86-based architectures (partly because we shouldn't be using floats in the kernel, and partly to avoid the need for stack alignment, see: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53383 ) UML does not do the same thing, which isn't in itself a problem, but does add to the list of differences between UML and "normal" x86 builds. In addition, there was a crash bug with LLVM < 15 / rustc < 1.65 when building with SSE, so disabling it fixes rust builds with earlier compiler versions, see: Rust-for-Linux/linux#881 Signed-off-by: David Gow <[email protected]> Reviewed-by: Sergio González Collado <[email protected]> Signed-off-by: Richard Weinberger <[email protected]> Signed-off-by: Sasha Levin <[email protected]>

commit 8849818 upstream. The kernel disables all SSE and similar FP/SIMD instructions on x86-based architectures (partly because we shouldn't be using floats in the kernel, and partly to avoid the need for stack alignment, see: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53383 ) UML does not do the same thing, which isn't in itself a problem, but does add to the list of differences between UML and "normal" x86 builds. In addition, there was a crash bug with LLVM < 15 / rustc < 1.65 when building with SSE, so disabling it fixes rust builds with earlier compiler versions, see: Rust-for-Linux/linux#881 Signed-off-by: David Gow <[email protected]> Reviewed-by: Sergio González Collado <[email protected]> Signed-off-by: Richard Weinberger <[email protected]> Signed-off-by: Greg Kroah-Hartman <[email protected]>

BugLink: https://bugs.launchpad.net/bugs/2025066 commit 8849818679478933dd1d9718741f4daa3f4e8b86 upstream. The kernel disables all SSE and similar FP/SIMD instructions on x86-based architectures (partly because we shouldn't be using floats in the kernel, and partly to avoid the need for stack alignment, see: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53383 ) UML does not do the same thing, which isn't in itself a problem, but does add to the list of differences between UML and "normal" x86 builds. In addition, there was a crash bug with LLVM < 15 / rustc < 1.65 when building with SSE, so disabling it fixes rust builds with earlier compiler versions, see: Rust-for-Linux/linux#881 Signed-off-by: David Gow <[email protected]> Reviewed-by: Sergio González Collado <[email protected]> Signed-off-by: Richard Weinberger <[email protected]> Signed-off-by: Greg Kroah-Hartman <[email protected]> Signed-off-by: Kamal Mostafa <[email protected]> Signed-off-by: Stefan Bader <[email protected]>

[ Upstream commit 8849818679478933dd1d9718741f4daa3f4e8b86 ] The kernel disables all SSE and similar FP/SIMD instructions on x86-based architectures (partly because we shouldn't be using floats in the kernel, and partly to avoid the need for stack alignment, see: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53383 ) UML does not do the same thing, which isn't in itself a problem, but does add to the list of differences between UML and "normal" x86 builds. In addition, there was a crash bug with LLVM < 15 / rustc < 1.65 when building with SSE, so disabling it fixes rust builds with earlier compiler versions, see: Rust-for-Linux/linux#881 Signed-off-by: David Gow <[email protected]> Reviewed-by: Sergio González Collado <[email protected]> Signed-off-by: Richard Weinberger <[email protected]> Signed-off-by: Sasha Levin <[email protected]>

bjorn3 reviewed Sep 12, 2022

View reviewed changes

scripts/generate_rust_target.rs Outdated Show resolved Hide resolved

sulix force-pushed the rust_uml_fix branch 2 times, most recently from 3015c66 to 3aa68a3 Compare September 12, 2022 15:37

sulix mentioned this pull request Sep 22, 2022

hermiticity issues with ARCH=um ClangBuiltLinux/linux#1715

Open

JoseExposito mentioned this pull request Dec 5, 2022

Add rust/.kunitconfig and document unit testing #935

Open

sulix force-pushed the rust_uml_fix branch from 3aa68a3 to ccb389f Compare December 8, 2022 06:50

sulix marked this pull request as ready for review December 8, 2022 07:20

bjorn3 approved these changes Dec 8, 2022

View reviewed changes

ojeda reviewed Dec 10, 2022

View reviewed changes

sulix force-pushed the rust_uml_fix branch from ccb389f to ad01833 Compare December 15, 2022 14:05

sulix added 2 commits December 17, 2022 11:48

rust: Use 'pie' relocation mode under UML

3b3136a

UML expects a position independent executable for some reason, so tell rustc to generate pie objects. Otherwise we get a bunch of relocations we can't deal with in libcore. Signed-off-by: David Gow <[email protected]>

sulix force-pushed the rust_uml_fix branch from ad01833 to 11c41c4 Compare December 17, 2022 03:49

ojeda merged commit 9c0096c into Rust-for-Linux:rust Feb 13, 2023

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Fix building under UML #881

Fix building under UML #881

sulix commented Sep 12, 2022

sulix commented Dec 8, 2022

bjorn3 left a comment

ojeda Dec 8, 2022

ojeda Dec 9, 2022 •

edited

Loading

sulix commented Dec 15, 2022

ojeda commented Dec 15, 2022

sulix commented Dec 17, 2022

sulix commented Feb 13, 2023

ojeda commented Feb 13, 2023

Fix building under UML #881

Fix building under UML #881

Conversation

sulix commented Sep 12, 2022

sulix commented Dec 8, 2022

bjorn3 left a comment

Choose a reason for hiding this comment

ojeda Dec 8, 2022

Choose a reason for hiding this comment

ojeda Dec 9, 2022 • edited Loading

Choose a reason for hiding this comment

sulix commented Dec 15, 2022

ojeda commented Dec 15, 2022

sulix commented Dec 17, 2022

sulix commented Feb 13, 2023

ojeda commented Feb 13, 2023

ojeda Dec 9, 2022 •

edited

Loading