Re: [PATCH v3 1/2] RISC-V: Define sys_riscv_flush_icache when SMP=n

[Palmer Dabbelt]

On Fri, 10 Aug 2018 11:47:15 PDT (-0700), linux@xxxxxxxxxxxx wrote:
> On Fri, Aug 10, 2018 at 11:27:37AM -0700, Palmer Dabbelt wrote:
> > On Fri, 10 Aug 2018 01:38:04 PDT (-0700), Christoph Hellwig wrote:
> > >On Thu, Aug 09, 2018 at 03:19:51PM -0700, Palmer Dabbelt wrote:
> > >>This would be necessary to make non-SMP builds work, but there is
> > >>another error in the implementation of our syscall linkage that actually
> > >>just causes sys_riscv_flush_icache to never build.  I've build tested
> > >>this on allnoconfig and allnoconfig+SMP=y, as well as defconfig like
> > >>normal.
> > >
> > >Would't it make sense to use COND_SYSCALL to stub out the syscall
> > >for !SMP builds?
> >
> > I'm not sure.  We can implement the syscall fine in !SMP, it's just that the
> > vDSO is expected to always eat these calls because in non-SMP mode you can
> > do a global fence.i by just doing a local fence.i (there's only one hart).
> >
> > The original rationale behind not having the syscall in non-SMP mode was to
> > limit the user ABI, but on looking again that seems like it's just a bit of
> > extra complexity that doesn't help anything.  It's already been demonstrated
>
> Doesn't this mean that some userspace code will only run if the kernel was
> compiled for SMP ? I always thought that was unacceptable.

Well, the officially sanctioned way to obtain this functionality is via a vDSO 
call.  On non-SMP systems it will never make the system call.  As a result we 
thought we'd keep it out of the ABI, but after looking again it seems yucky to 
do so.  Here's the vDSO entry, for reference:

   ENTRY(__vdso_flush_icache)
           .cfi_startproc
   #ifdef CONFIG_SMP
           li a7, __NR_riscv_flush_icache
           ecall
   #else
           fence.i
           li a0, 0
   #endif
           ret
           .cfi_endproc
   ENDPROC(__vdso_flush_icache)

Note that glibc has a fallback to make the system call if it can't find the 
vDSO entry, but then doesn't have a secondary fallback to emit a local fence.i 
if the system call doesn't exist.  It seems easier to fix the kernel to always 
provide the syscall and just call it a bug.