[RFC2 nowrap: PATCH v7 00/18] ILP32 for ARM64

[catalin.marinas@arm.com (Catalin Marinas)]

On Wed, Aug 17, 2016 at 04:32:23PM +0200, Dr. Philipp Tomsich wrote:
> On 17 Aug 2016, at 16:29, Catalin Marinas <catalin.marinas@arm.com> wrote:
> > On Wed, Aug 17, 2016 at 02:54:59PM +0200, Dr. Philipp Tomsich wrote:
> >> On 17 Aug 2016, at 14:48, Yury Norov <ynorov@caviumnetworks.com> wrote:
> >>> On Wed, Aug 17, 2016 at 02:28:50PM +0200, Alexander Graf wrote:
> >>>> On 17 Aug 2016, at 13:46, Yury Norov <ynorov@caviumnetworks.com> wrote:
> >>>>> This series enables aarch64 with ilp32 mode, and as supporting work,
> >>>>> introduces ARCH_32BIT_OFF_T configuration option that is enabled for
> >>>>> existing 32-bit architectures but disabled for new arches (so 64-bit
> >>>>> off_t is is used by new userspace).
> >>>>> 
> >>>>> This version is based on kernel v4.8-rc2.
> >>>>> It works with glibc-2.23, and tested with LTP.
> >>>>> 
> >>>>> This is RFC because there is still no solid understanding what type of registers
> >>>>> top-halves delousing we prefer. In this patchset, w0-w7 are cleared for each
> >>>>> syscall in assembler entry. The alternative approach is in introducing compat
> >>>>> wrappers which is little faster for natively routed syscalls (~2.6% for syscall
> >>>>> with no payload) but much more complicated.
> >>>> 
> >>>> So you?re saying there are 2 options:
> >>>> 
> >>>> 1) easy to get right, slightly slower, same ABI to user space as 2
> >>>> 2) harder to get right, minor performance benefit
> >>> 
> >>> No, ABI is little different. If 1) we pass off_t in a pair to syscalls,
> >>> if 2) - in a single register. So if 1, we 'd take some wrappers from aarch32.
> >>> See patch 12 here.
> >> 
> >> From our experience with ILP32, I?d prefer to have off_t (and similar)
> >> in a single register whenever possible (i.e. option #2).  It feels
> >> more natural to use the full 64bit registers whenever possible, as
> >> ILP32 on ARMv8 should really be understood as a 64bit ABI with a 32bit
> >> memory model.
> > 
> > I think we are well past the point where we considered ILP32 a 64-bit
> > ABI. It would have been nice but we decided that breaking POSIX
> > compatibility is a bad idea, so we went back (again) to a 32-bit ABI for
> > ILP32. While there are 64-bit arguments that, at a first look, would
> > make sense to be passed in 64-bit registers, the kernel maintenance cost
> > is significant with changes to generic files.
> > 
> > Allowing 64-bit wide registers at the ILP32 syscall interface means that
> > the kernel would have to zero/sign-extend the upper half of the 32-bit
> > arguments for the cases where they are passed directly to a native
> > syscall that expects a 64-bit argument. This (a) adds a significant
> > number of wrappers to the generic code together additional annotations
> > to the generic unistd.h and (b) it adds a small overhead to the AArch32
> > (compat) ABI since it doesn't need such generic wrapping (the upper half
> > of 64-bit registers is guaranteed to be zero/preserved by the
> > architecture when coming from the AArch32 mode).
> 
> Yes, I remember the discussions and just wanted to put option #2 in
> context again.

I don't particularly like splitting 64-bit arguments in two 32-bit
values either but I don't see a better alternative. To keep this
mostly in the arch code we would need an additional table of syscall
wrappers where the majority just use the default zero-extend everything
with a few specific wrappers where we pass 64-bit arguments. Or we could
set an extra bit in the syscall number for those syscalls that need
special wrapping and avoid zero-extending. But neither of these look any
nicer (well, maybe only from the user-space perspective).

> Everything points to just going with the pair-of-registers and getting
> this merged quickly then, I suppose.

I will refrain from commenting on how quickly we merge this ;) (it may
be seen as binding by some).

-- 
Catalin