Re: [PATCH V4 6/8] riscv: Support HAVE_IRQ_EXIT_ON_IRQ_STACK

[Guo Ren]

On Fri, Sep 9, 2022 at 3:30 PM David Laight <David.Laight@xxxxxxxxxx> wrote:
>
> From: Sebastian Andrzej Siewior
> > Sent: 08 September 2022 17:08
> >
> > On 2022-09-07 22:25:04 [-0400], guoren@xxxxxxxxxx wrote:
> > > diff --git a/arch/riscv/Kconfig b/arch/riscv/Kconfig
> > > index a07bb3b73b5b..a8a12b4ba1a9 100644
> > > --- a/arch/riscv/Kconfig
> > > +++ b/arch/riscv/Kconfig
> > > @@ -433,6 +433,14 @@ config FPU
> > >
> > >       If you don't know what to do here, say Y.
> > >
> > > +config IRQ_STACKS
> > > +   bool "Independent irq stacks"
> > > +   default y
> > > +   select HAVE_IRQ_EXIT_ON_IRQ_STACK
> > > +   help
> > > +     Add independent irq stacks for percpu to prevent kernel stack overflows.
> > > +     We may save some memory footprint by disabling IRQ_STACKS.
> >
> > Do you really think that it is needed to save memory here? Avoiding
> > stack overflows in deep call chains is probably more important than
> > saving ~8KiB per CPU.
Original riscv is !IRQ_STACKS, I just give a config to make it back.
So I would add a CONFIG_EXPERT in the next version.

Actually, I have a similar opinion to you, IRQ_STACKS should be force
enabled. But as a new feature, we should give users a choice - use or
not.

>
> Particularly if a 64bit build is using small stacks.
>
> Without static analysis of actual call chain depth it is
> really difficult to trim the stack size.
>
> I'd bet (a few beers) that the deepest stack use in inside
> the console print code form a printk() (eg warn_on_once)
> in an obscure error path somewhere.
> This won't be hit during any normal testing.
That means stack overflow would be hidden a lot. But we could enable
VMAP_STACK & STACK_LEAK [1].

[1]: https://lore.kernel.org/lkml/20220907014809.919979-1-guoren@xxxxxxxxxx/

>
> I think that the analysis objtool does is getting close
> to be able to generate the raw data that can be used for
> static stack depth analysis.
> You need the 'CFI' constants for indirect calls and
> some assumptions about depth of recursive calls.
> But apart from that the code to process the raw output
> isn't that complex.
>
> A nice task for someone with some spare time.
>
>         David
>
> -
> Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
> Registration No: 1397386 (Wales)



-- 
Best Regards
 Guo Ren