On Wed, 7 Jul 2021 at 12:19, Marco Elver <elver@xxxxxxxxxx> wrote: > On Sun, Jul 04, 2021 at 10:27PM +0200, ojeda@xxxxxxxxxx wrote: > > From: Miguel Ojeda <ojeda@xxxxxxxxxx> > > > > This source file contains forwarders to C macros and inlined > > functions. > > What is the story with Rust and LTO? Intuitively, I would expect Rust > code to only perform optimally if the kernel is built with LTO > (currently only supported via Clang). I'll answer my own question: it looks like Linux Rust code currently does _not_ generate LLVM-LTO compatible object files, but only native object files (which still link fine if LTO is enabled, but doesn't permit the optimizations below we'd want). rustc already supports playing nicely with LLVM LTO via `-C linker-plugin-lto`: https://doc.rust-lang.org/rustc/linker-plugin-lto.html So, hopefully it should only require kernel work to make it play nicely with CONFIG_LTO_CLANG. > Because if calls to every one of these helpers are real calls, I would > expect performance to be pretty poor. There's probably a reason these > are macros or inlinable functions. > > I would almost go so far and suggest that CONFIG_RUST be modified as > follows: > > --- a/init/Kconfig > +++ b/init/Kconfig > @@ -2028,6 +2028,7 @@ config RUST > depends on HAS_RUST > depends on !COMPILE_TEST > depends on !MODVERSIONS > + depends on LTO || EXPERT > default n > help > Enables Rust support in the kernel. > > [ I'm sure there are configs that don't yet work with LTO, but could be > useful to enable for debugging or testing purposes, and therefore would > make it conditional on CONFIG_EXPERT as well. ] > > [...] > > +unsigned long rust_helper_copy_from_user(void *to, const void __user *from, unsigned long n) > > +{ > > + return copy_from_user(to, from, n); > > +} > > + > [...] > > From some local tests, it looks like simply attaching > __attribute__((always_inline)) will do what one would expect when > compiling with Clang LTO (I checked -flto=thin). > > If you confirm this also works across C and Rust TUs when enabling LTO, > I would then suggested adding __attribute__((always_inline)) to all > these helpers. > > Thanks, > -- Marco
