Re: [PATCH] rust: kasan/kbuild: fix missing flags on first build

[Conor Dooley <conor@kernel.org>]

[-- Attachment #1: Type: text/plain, Size: 6781 bytes --]

On Tue, Sep 02, 2025 at 08:29:29AM +0000, Alice Ryhl wrote:
> On Mon, Sep 01, 2025 at 06:45:54PM +0100, Conor Dooley wrote:
> > Yo,
> > 
> > On Wed, Apr 09, 2025 at 12:03:11AM +0200, Miguel Ojeda wrote:
> > > If KASAN is enabled, and one runs in a clean repository e.g.:
> > > 
> > >     make LLVM=1 prepare
> > >     make LLVM=1 prepare
> > > 
> > > Then the Rust code gets rebuilt, which should not happen.
> > > 
> > > The reason is some of the LLVM KASAN `rustc` flags are added in the
> > > second run:
> > > 
> > >     -Cllvm-args=-asan-instrumentation-with-call-threshold=10000
> > >     -Cllvm-args=-asan-stack=0
> > >     -Cllvm-args=-asan-globals=1
> > >     -Cllvm-args=-asan-kernel-mem-intrinsic-prefix=1
> > > 
> > > Further runs do not rebuild Rust because the flags do not change anymore.
> > > 
> > > Rebuilding like that in the second run is bad, even if this just happens
> > > with KASAN enabled, but missing flags in the first one is even worse.
> > > 
> > > The root issue is that we pass, for some architectures and for the moment,
> > > a generated `target.json` file. That file is not ready by the time `rustc`
> > > gets called for the flag test, and thus the flag test fails just because
> > > the file is not available, e.g.:
> > > 
> > >     $ ... --target=./scripts/target.json ... -Cllvm-args=...
> > >     error: target file "./scripts/target.json" does not exist
> > > 
> > > There are a few approaches we could take here to solve this. For instance,
> > > we could ensure that every time that the config is rebuilt, we regenerate
> > > the file and recompute the flags. Or we could use the LLVM version to
> > > check for these flags, instead of testing the flag (which may have other
> > > advantages, such as allowing us to detect renames on the LLVM side).
> > > 
> > > However, it may be easier than that: `rustc` is aware of the `-Cllvm-args`
> > > regardless of the `--target` (e.g. I checked that the list printed
> > > is the same, plus that I can check for these flags even if I pass
> > > a completely unrelated target), and thus we can just eliminate the
> > > dependency completely.
> > > 
> > > Thus filter out the target.
> > 
> > 
> > 
> > 
> > > This does mean that `rustc-option` cannot be used to test a flag that
> > > requires the right target, but we don't have other users yet, it is a
> > > minimal change and we want to get rid of custom targets in the future.
> > 
> > Hmm, while this might be true, I think it should not actually have been
> > true. Commit ca627e636551e ("rust: cfi: add support for CFI_CLANG with Rust")
> > added a cc-option check to the rust kconfig symbol, checking if the c
> > compiler supports the integer normalisations stuff:
> > 	depends on !CFI_CLANG || RUSTC_VERSION >= 107900 && $(cc-option,-fsanitize=kcfi -fsanitize-cfi-icall-experimental-normalize-integers)
> > and also sets the relevant options in the makefile:
> > 	ifdef CONFIG_RUST
> > 	       # Always pass -Zsanitizer-cfi-normalize-integers as CONFIG_RUST selects
> > 	       # CONFIG_CFI_ICALL_NORMALIZE_INTEGERS.
> > 	       RUSTC_FLAGS_CFI   := -Zsanitizer=kcfi -Zsanitizer-cfi-normalize-integers
> > 	       KBUILD_RUSTFLAGS += $(RUSTC_FLAGS_CFI)
> > 	       export RUSTC_FLAGS_CFI
> > 	endif
> > but it should also have added a rustc-option check as, unfortunately,
> > support for kcfi in rustc is target specific. This results in build
> > breakages where the arch supports CFI_CLANG and RUST, but the target in
> > use does not have the kcfi flag set.
> > I attempted to fix this by adding:
> > 	diff --git a/arch/Kconfig b/arch/Kconfig
> > 	index d1b4ffd6e0856..235709fb75152 100644
> > 	--- a/arch/Kconfig
> > 	+++ b/arch/Kconfig
> > 	@@ -916,6 +916,7 @@ config HAVE_CFI_ICALL_NORMALIZE_INTEGERS_CLANG
> > 	 config HAVE_CFI_ICALL_NORMALIZE_INTEGERS_RUSTC
> > 	        def_bool y
> > 	        depends on HAVE_CFI_ICALL_NORMALIZE_INTEGERS_CLANG
> > 	+       depends on $(rustc-option,-C panic=abort -Zsanitizer=kcfi -Zsanitizer-cfi-normalize-integers)
> > 	        depends on RUSTC_VERSION >= 107900
> > 	        # With GCOV/KASAN we need this fix: https://github.com/rust-lang/rust/pull/129373
> > 	        depends on (RUSTC_LLVM_VERSION >= 190103 && RUSTC_VERSION >= 108200) || \
> > but of course this does not work for cross compilation, as you're
> > stripping the target information out and so the check passes on my host
> > even though my intended
> > RUSTC_BOOTSTRAP=1 rustc -C panic=abort -Zsanitizer=kcfi -Zsanitizer-cfi-normalize-integers -Ctarget-cpu=generic-rv64 --target=riscv64imac-unknown-none-elf
> > is a failure.
> > 
> > I dunno too much about rustc itself, but I suspect that adding kcfi to
> > the target there is a "free" win, but that'll take time to trickle down
> > and the minimum version rustc version for the kernel isn't going to have
> > that.
> > 
> > I'm not really sure what your target.json suggestion below is, so just
> > reporting so that someone that understands the alternative solutions can
> > fix this.
> 
> Probably right now we have to do this cfg by
> 
> 	depends on CONFIG_ARM

It's valid on x86 too, right?

> 
> to prevent riscv if rustc has the missing setting
> set on riscv. Once we add it to riscv, we change it to
> 
> 	depends on CONFIG_ARM || (RUSTC_VERSION >= ??? || CONFIG_RISCV)

I kinda shied away from something like this since there was already a
cc-option on the other half and checking different versions per arch
becomes a mess - but yeah it kinda is a no-brainer to do it here when
rustc-option is kinda broken.

I guess the temporary fix is then:

config HAVE_CFI_ICALL_NORMALIZE_INTEGERS_RUSTC
	def_bool y
	depends on HAVE_CFI_ICALL_NORMALIZE_INTEGERS_CLANG
	depends on ARM64 || x86_64
	depends on RUSTC_VERSION >= 107900
	# With GCOV/KASAN we need this fix: https://github.com/rust-lang/rust/pull/129373
	depends on (RUSTC_LLVM_VERSION >= 190103 && RUSTC_VERSION >= 108200) || \
		(!GCOV_KERNEL && !KASAN_GENERIC && !KASAN_SW_TAGS)

because there's no 32-bit target with SanitizerSet::KCFI in rustc either
AFAICT. Then later on it'd become more like:

config HAVE_CFI_ICALL_NORMALIZE_INTEGERS_RUSTC
	def_bool y
	depends on HAVE_CFI_ICALL_NORMALIZE_INTEGERS_CLANG
	depends on RISCV || ((ARM64 || x86_64) && RUSTC_VERSION >= 107900)
	depends on (ARM64 || x86_64) || (RISCV && RUSTC_VERSION >= 999999)
	# With GCOV/KASAN we need this fix: https://github.com/rust-lang/rust/pull/129373
	depends on (RUSTC_LLVM_VERSION >= 190103 && RUSTC_VERSION >= 108200) || \
		(!GCOV_KERNEL && !KASAN_GENERIC && !KASAN_SW_TAGS)

but that exact sort of mess is what becomes unwieldy fast since that
doesn't even cover 32-bit arm.


[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 228 bytes --]