Re: [PATCH 09/21] ARC: Linux ABI

[Vineet Gupta <vineet dot gupta1 at synopsys dot com>]

On 12/19/18 12:36 PM, Joseph Myers wrote:
> On Wed, 19 Dec 2018, Vineet Gupta wrote:
> 
>> So perhaps we keep the ARC header above, with matching layout, while 
>> switching to generic sigaction() still. OK ?
> 
> If the different userspace layout is better, it should be better for all 
> future ports as well.
> 
> Duplicating all the SA_* constants in architecture-specific headers, even 
> when those architectures don't actually have different values of those 
> constants, is bad, especially since there is logic about which __USE_* 
> macros get which SA_* defined which then needs to be duplicated.

Totally agree.

> So what this suggests to me is a preliminary patch series to unify 
> bits/sigaction.h headers - there might be one header for SA_* constants 
> (architecture-specific variants only when needed), one for SIG_* constants 
> (architecture-specific variants only when needed), and one for the ways in 
> which the layout of the userspace struct sigaction can vary (defining 
> macros that the unified bits/sigaction.h then uses to control the layout - 
> cf. the siginfo unification a while back).

Actually this can be simplified. Given the asm-generic syscall ABI, we are anyhow
locked into SA_* or SIG_* and kernel side sigaction, the only variant is user
facing struct sigaction. I suppose we could leave it at that given that only 1
syscall sigaction is affected. The bigger bang for the buck will the reduction of
sizeof(sigset_t) - purely from amount of memory any sig* call needs to
read/write/copy.

I'll follow Adhemerval's suggestion of arch-specific __sigset_t.h which seems
doable w/o much pain.

> Initially, the default version of that last header would correspond to the 
> existing sysdeps/unix/sysv/linux/bits/sigaction.h header - however, if you 
> get consensus for a different layout as the default for new architectures, 
> the default would change and all existing architectures that use the 
> current default layout would get copies of the new bits/ header configured 
> to keep the old layout.  (Variants are possible to reduce the number of 
> architectures needing architecture-specific copies as a result of the 
> changed default.  For example, you could have the new default only in 
> sysdeps/unix/sysv/linux/generic not sysdeps/unix/sysv/linux/ and then only 
> existing asm-generic architectures (aarch64 nios2 riscv) would need to 
> override back to the old default layout.)

Yes this will require a needless churn all over, just for 1 syscall.