Re: rcutorture initrd/nolibc build on ARMv8?

[Willy Tarreau <w@1wt.eu>]

On Wed, Jan 20, 2021 at 02:54:47PM +0000, Mark Rutland wrote:
> On Wed, Jan 20, 2021 at 03:25:00PM +0100, Willy Tarreau wrote:
> > On Wed, Jan 20, 2021 at 01:45:11PM +0000, Mark Rutland wrote:
> > > > Ah that's very interesting indeed because actually these ones should
> > > > only be used when __NR_dup3 or __NR_clone are not defined. Thus I wanted
> > > > to check the definitions that were reported in your error output but
> > > > actually what was needed was to figure whether the correct ones were
> > > > present, and they are, here on my machine (and yes I agree that in this
> > > > case the dup2/fork above are bofus):
> > > 
> > > The issue is that even if a function is unused, the compiler still has
> > > to parse and compile the code, so where __NR_dup2 is not defined, we'll
> > > get a build error for:
> > > 
> > > static __attribute__((unused))
> > > int sys_dup2(int old, int new)
> > > {
> > >        return my_syscall2(__NR_dup2, old, new);
> > > }
> > 
> > For sure but this is supposed to be used only when __NR_dup3 is not
> > defined. Ah now I understand where my mistake is: after it built
> > successfully for me I inspected the most recent tree which has these
> > in place. Sorry for being stupid!
> > 
> > In my local tree it's defined like this:
> > 
> >  static __attribute__((unused))
> >  int sys_dup2(int old, int new)
> >  {
> > #ifdef __NR_dup3
> >        return my_syscall3(__NR_dup3, old, new, 0);
> > #else
> >        return my_syscall2(__NR_dup2, old, new);
> > #endif
> >  }
> 
> Ah, much better!
> 
> For robustness, I think it would be worth doing:
> 
> static __attribute__((unused))
> int sys_dup2(int old, int new)
> {
> #if defined(__NR_dup3)
> 	return my_syscall3(__NR_dup3, old, new, 0);
> #elif defined(__NR_dup2)
> 	return my_syscall2(__NR_dup2, old, new);
> #else
> #error Cannot implement dup2
> #endif
> }

Yep I'm fine with this. We'll probably adjust between return -ENOSYS
or #error depending on the "criticity" of the syscalls. E.g. we can
miss poll() for plenty of init stuff band be done with ENOSYS but
missing dup is quickly going to cause trouble.

> ... and getting rid of the ARCH_WANT_* definitions (which are never
> legitimate for userspace to set). That way, there's no risk that we
> accidentally use a bogus syscall number in future. Where the kernel does
> implement a syscall, it will have done whatever is necessary to expose
> the corresponding __NR_<syscall> to userspace without userspace needing
> to define anything.

I'm really willing to get rid of that crap, I hated it, I vaguely
remember having gone through layers of glibc indirections when using
the pre-processed asm/* and found this to be the only way to expose
some of them. The fact that it's not needed for you is pretty much
encouraging. I'll just retry on an older libc I've used a lot (2.18)
to make sure I didn't overlook anything.

> > However I'm interested in knowing if the latest version fixes everything
> > for you or not :
> > 
> >   https://raw.githubusercontent.com/wtarreau/nolibc/master/nolibc.h
> 
> I replaced my in-tree copy with that, and it built without issues, and
> the tests ran correctly.

Perfect, thanks for the quick feedback!

> > OK thanks! I will retry here without setting those. I'm pretty sure I
> > needed these ones to find the __NR_* values but it's possible that it
> > was before I had the alternate ones and that these are finally not
> > nedeed at all (which I would prefer as these are ugly).
> 
> Great! I reckon they're not needed at all so long as usage of each
> __NR_* is suitably guarded (such as above).
> 
> If you do spot issues when removing ARCH_WANT_*, I'm happy to take a
> look, and/or to test patches handling any fallout.

Many thanks. I'm pretty confident that it's OK without them based on
your feedback but if need I'll ask you for some advices.

cheers,
Willy