Re: [PATCHv3 00/24] ILP32 support in ARM64

[Rich Felker <dalias at libc dot org>]

On Mon, Feb 16, 2015 at 03:40:54PM +0100, Arnd Bergmann wrote:
> On Friday 13 February 2015 13:37:07 Rich Felker wrote:
> > On Fri, Feb 13, 2015 at 05:33:46PM +0000, Catalin Marinas wrote:
> > > > > > The data structure definition is a little bit fragile, as it depends on
> > > > > > user space not using the __BIT_ENDIAN symbol in a conflicting way. So
> > > > > > far we have managed to keep that outside of general purpose headers, but
> > > > > > it should at least blow up in an obvious way if it does, rather than
> > > > > > breaking silently.
> > > > > > 
> > > > > > I still think it's more practical to keep the zeroing in user space though.
> > > > > > In that case, we keep defining __kernel_timespec64 with a 'typedef long
> > > > > > long __kernel_snseconds_t', and it's up to the libc to either use
> > > > > > __kernel_timespec64 as its timespec, or to define a C11-compliant
> > > > > > timespec itself and zero out the bits before passing the data to the kernel.
> > > > > 
> > > > > The problem with doing this in user space is syscall(2). If we don't
> > > > > allow it, then it's fine to do the padding in libc.
> > > > 
> > > > It's already the case that callers have to tiptoe around syscall(2)
> > > > usage on a per-arch basis for silly things like the convention for
> > > > passing 64-bit arguments on 32-bit archs, different arg orders to work
> > > > around 64-bit alignment and issues with too many args, and various
> > > > legacy issues.
> 
> Right. If one wants to use syscall(), they have to know exactly what the
> kernel's calling conventions are, including knowing what the timespec
> definition looks like, which could have a different size and padding
> compared to the user space one.
> 
> > > I think there is another problem with sign-extending tv_nsec in libc.
> > > The prototype for functions like clock_settime(2) take a const struct
> > > timespec *. There isn't anything to prevent such structure being in a
> > > read-only section, even though it is unlikely. So libc would have to
> > > duplicate the structure rather than just sign-extending tv_nsec in
> > > place.
> 
> Do we actually need sign-extend, or does zero-extend have the exact
> same effect? For all I can tell, all invalid nanoseconds values
> remain invalid, and the accepted values are unchanged regardless
> of which type extension gets used.

I think it matters for futimensat which has some special negative
codes you can store in tv_nsec, but perhaps there's an easy trick to
distinguish them even with zero extending.

> > Yes, we already have to do this for x32 in musl. I'd rather not have
> > to do the same for aarch64-ILP32.
> 
> This would of course be solved by using a 64-bit __kernel_snseconds_t
> or snseconds_t, and I suspect other libc implementations would just do
> that, when they are less strict about posix/c11 compliance compared
> to musl.

I think they would be more strict if this were for a target that
actually sees use and they were getting bug reports from C programmers
annoyed that their code was not working correctly or not even
compiling. AFAIK there are no distros based on x32 now and it's
something of an alternate model on x86_64 distros that some people are
playing around with.

> If you don't mind the (slight) distraction, can you describe what your
> plans are for handling 64-bit time_t on the existing 32-bit ABIs?
> I'm involved in both the efforts to do that and the ilp32 code on
> ARM, so it would be good for me to understand your plans for musl to
> get the bigger picture. Specifically, which of these do you plan
> to support (if you know already):

It largely depends on if there's demand. If we have users who want to
run 32-bit systems with an ABI that will survive Y2038, it will be
supported, but as a new ABI for these targets. This will likely allow
fixing other ABI issues at the same time -- for example, on i386 I
would probably switch to mandating SSE2 for floating point, and
possibly using regparm everywhere. There are a couple of different
ways it could be done though:

1. On a per-arch basis, defining a new ABI variant for the arch.

2. With a new abstraction at the syscall boundary to get rid of all
kernel-arch-specific structures in userspace and redefine all types to
have plenty of room for growth.

In regards to your specific questions about ways it could be done:

> - using 64-bit time_t on future arm32/i386/... kernels
> - using 64-bit time_t on existing arm32/i386/... kernels with native
>   32-bit time_t

If the former is supported, I would think we'd want to support the
latter too. An ABI that only works on very-new kernels is very
restrictive in who can use it. Kernel support hardly matters (until
Y2038 actually arrives); the point of 64-bit time_t is to have an ABI
that's _ready_ for it so existing binaries can keep working.

> - using 32-bit time_t on future architectures that only support 64-bit
>   time_t in the kernel

Definitely will not be supported. Introducing a new ABI with 32-bit
time_t is a huge mistake, and the only reason it's been done for some
of the new targets musl supports is because the kernel does it, and
working around a mismatch between kernel and user time_t is a huge
problem -- all sorts of things, including for example struct stat,
depend on the time_t definition, and if you're going to allow mismatch
with kernel you might as well go ahead and have a full translation
layer for kernel structs like this.

> - running existing binaries with 32-bit time_t on a library with 64-bit
>   time_t support, using symbol versioning

Symbol versions don't solve any problem, and they mask dangerous bugs,
so no. The problem is that a symbol version is only able to represent
a single interface boundary (between a caller and libc), not all the
other interface boundaries between third-party libraries. If code
compiled for 32-bit time_t calls into code that uses 64-bit time_t
with a time_t* argument and the callee writes back a result, it's
corrupted the caller's memory. Symbol versions have no way to diagnose
this.

They're also bound at ld-time, whereas the choice of needed version
depends on compile-time (which definitions were used in the header the
code was compiled against).

> - compiling new code with 32-bit time_t against a library that supports
>   both 32-bit and 64-bit time_t at runtime.

No; see above.

> - building a libc for existing architectures but without support for
>   running existing 32-bit time_t applications.

Yes; this would be the way a new ABI would always work. But since musl
inherently supports multi-arch (each arch variant has its own
PT_INTERP name and library path config) you can easily run both types
of binaries on the same system. They just need completely separate
library ecosystems. This is the only way I know to prevent the
dangerous issues that arise with other [non-]solutions like symbol
versioning or feature test macros (as in -D_FILE_OFFSET_BITS=64) for
the problem.

Rich