Re: [PATCH v3 3/5] RISC-V: Align the .init.text section

[Palmer Dabbelt]

On Tue, 15 Dec 2020 22:02:54 PST (-0800), Palmer Dabbelt wrote:
> On Wed, 04 Nov 2020 16:04:37 PST (-0800), Atish Patra wrote:
> > In order to improve kernel text protection, we need separate .init.text/
> > .init.data/.text in separate sections. However, RISC-V linker relaxation
> > code is not aware of any alignment between sections. As a result, it may
> > relax any RISCV_CALL relocations between sections to JAL without realizing
> > that an inter section alignment may move the address farther. That may
> > lead to a relocation truncated fit error. However, linker relaxation code
> > is aware of the individual section alignments.
> >
> > The detailed discussion on this issue can be found here.
> > https://github.com/riscv/riscv-gnu-toolchain/issues/738
> >
> > Keep the .init.text section aligned so that linker relaxation will take
> > that as a hint while relaxing inter section calls.
> > Here are the code size changes for each section because of this change.
> >
> > section         change in size (in bytes)
> >   .head.text      +4
> >   .text           +40
> >   .init.text      +6530
> >   .exit.text      +84
> >
> > The only significant increase in size happened for .init.text because
> > all intra relocations also use 2MB alignment.
> >
> > Suggested-by: Jim Wilson <jimw@xxxxxxxxxx>
> > Signed-off-by: Atish Patra <atish.patra@xxxxxxx>
> > ---
> >  arch/riscv/kernel/vmlinux.lds.S | 8 +++++++-
> >  1 file changed, 7 insertions(+), 1 deletion(-)
> >
> > diff --git a/arch/riscv/kernel/vmlinux.lds.S b/arch/riscv/kernel/vmlinux.lds.S
> > index 3ffbd6cbdb86..cacd7898ba7f 100644
> > --- a/arch/riscv/kernel/vmlinux.lds.S
> > +++ b/arch/riscv/kernel/vmlinux.lds.S
> > @@ -30,7 +30,13 @@ SECTIONS
> >  	. = ALIGN(PAGE_SIZE);
> >
> >  	__init_begin = .;
> > -	INIT_TEXT_SECTION(PAGE_SIZE)
> > +	__init_text_begin = .;
> > +	.init.text : AT(ADDR(.init.text) - LOAD_OFFSET) ALIGN(SECTION_ALIGN) { \
> > +		_sinittext = .;						\
> > +		INIT_TEXT						\
> > +		_einittext = .;						\
> > +	}
> > +
> >  	. = ALIGN(8);
> >  	__soc_early_init_table : {
> >  		__soc_early_init_table_start = .;
>
> Not sure what's going on here (or why I wasn't catching it earlier), but this
> is breaking boot on one of my test configs.  I'm not getting any Linux boot
> spew, so it's something fairly early.  I'm running defconfig with
>
>     CONFIG_PREEMPT=y
>     CONFIG_DEBUG_PREEMPT=y
>     CONFIG_PROVE_LOCKING=y
>
> It looks like that's been throwing a bunch of warnings for a while, but it did
> at least used to boot.  No idea what PREEMPT would have to do with this, and
> the other two don't generally trigger issues that early in boot (or at least,
> trigger halts that early in boot).
>
> There's a bunch of other stuff that depends on this that's on for-next so I
> don't want to just drop it, but I also don't want to break something.  I'm just
> running QEMU's virt board.
>
> I'll take a look again tomorrow night, but if anyone has some time to look
> that'd be great!

Looks like this breaks on QEMU 5.0.0 but works on 5.2.0.  I guess technically
that means could be considered a regression, but as we don't really have any
scheme for which old versions of QEMU we support it's not absolute.  I'd
usually err on the side of keeping support for older platforms, but in this
case it's probably just not worth the time so I'm going to just ignore it.