[PATCH v4 00/24] ILP32 for ARM64

[Arnd Bergmann]

On Tuesday 14 April 2015 13:50:21 Dr. Philipp Tomsich wrote:
> 
> > On 14 Apr 2015, at 13:14, Arnd Bergmann <arnd at arndb.de> wrote:
> > 
> > On Tuesday 14 April 2015 10:45:43 Pinski, Andrew wrote:
> >>> On Apr 14, 2015, at 3:08 AM, Arnd Bergmann <arnd at arndb.de> wrote:
> >>> 
> >>>> On Tuesday 14 April 2015 11:33:13 Dr.  Philipp Tomsich wrote:
> >>>> Arnd,
> >>>> 
> >>>> After getting a good night’s sleep, the “reuse the existing system call table” comment
> >>>> makes a little more sense as I construe it as having just one merged system call table
> >>>> for both LP64 and ILP32 and handling the differences through a different system call
> >>>> numbering in unistd.h towards LP64 and ILP32 processes.
> >>>> 
> >>>> If this is the intended implementation, I am not fully sold on the benefit: having a private
> >>>> copy of unistd.h for ARM64 seems to be a less readable and less maintenance-friendly
> >>>> solution to having separate tables.
> >>>> 
> >>>> We’re open to input on this and—if merging the system call tables is the consensus—
> >>>> would like to get the change underway as soon as possible.
> >>> 
> >>> There are multiple ways of doing this:
> >>> 
> >>> a) separate syscall table for arm64: as you say, this is the current approach,
> >>>  and I'd like to avoid that too
> >>> b) add syscalls for ilp32 as additional numbers in the normal lp64 version of
> >>>  asm-generic/unistd.h, and share the binary tables between ilp32 and lp64
> >>>  on aarch64
> >>> c) change asm-generic/unistd.h to generate three possible tables: instead of
> >>>  just native (lp64 or ilp32 depending on the arch), compat (support for
> >>>  existing ilp32 binaries on some architectures, there would also be a
> >>>  "modern" ilp32 variant that is a mix of the two, as your table today
> >>> d) don't use the asm-generic/unistd.h table for aarch64-ilp32 at all, but instead
> >>>  reuse the table from arch/arm64/include/asm/unistd32.h
> >>> 
> >>> I think you are referring to approach b) or c) above, but my preferred one
> >>> would actually be d).
> >> 
> >> D is the worst of all 4 options in my mind. The reason is when a new syscall is
> >> added, then you have to update that file too.
> > 
> > I don't know what the miscommunication is here, but the advantage of d is
> > specifically that it is /less/ work to maintain: With the current approach,
> > each new syscall that gets added needs to be checked to see if the normal
> > aarch64 version works or if it needs another wrapper, while with d) we
> > get the update for free, because we follow exactly what aarch32 is doing.
> 
> I must agree with Andrew, that (d) seems like a bad fit for ILP32… after all the
> ILP32 (ELF) ABI specifies that 64bit values are to be passed in a single register,
> but the unistd32.h assumes a 32bit kernel that receives 64bit arguments split
> over two registers (i.e. the 64-variants of the various system calls, such as
> ftruncate64).
> 
> I strongly prefer (b) as this satisfies the largest number of requirements:
>  (i) it will be a single system call table for LP64 and ILP32
>  (ii) it’s easy to be make use of the 64bit capable system calls
>  (iii) it fits with the relationship of ILP32 to LP64 in the ILP32 ELF ABI definition

Ok, I see. I'd still like to hear other opinions on the matter, and my
preference remains D. Most importantly, I'd like Catalin and Will to
comment on this.

For completeness, there is yet another option, which would be to use the
exact system call table from arm64 and do all the emulation in user space
rather than the kernel. This would however be the least compatible with
existing source code, so you probably don't want to do that.

> As we don’t support ILP32 without LP64, the resulting implementation will
> be even simpler (i.e. no need to duplicate the entire table) than the n32 
> implementation on MIPS...

Right. Andrew didn't like the idea that the syscall numbers are slightly
different in B, but I think that is not a significant downside, as the kernel
header would be done in a way to report the correct __NR_* macros here.

> >> Also d is worse than the rest as
> >> you no longer default to 64bit off_t which is not a good thing. 
> > 
> > That decision is up to the libc implementation, just as it is for the existing
> > aarch32 libc. The kernel just offers both versions and the libc can pick
> > one, or use the _LARGEFILE64_SOURCE hack that glibc has to also implement
> > both. It would probably be reasonable to use 64-bit off_t only for a libc
> > and ignore the old calls.
> 
> The glibc implementation, as we have it today, always uses the 64bit system call
> and performs overflow-checking on results (on ILP32, we can’t perform overflow
> checking on arguments, as the callee needs to sign-extend).
> 
> In other words: glibc uses the LP64 system calls and handles any pre- and 
> post-processing in the system-call wrappers.

Hmm, that sounds like much more work you already do in glibc than
you'd need to split up the 64-bit arguments for the eight syscalls
that pass an loff_t in a register.

> >> B is just as bad and goes against using the generic syscall numbers. 
> > 
> > How so? The newly introduce syscalls then would be the generic ones.
> 
> By applying the “strace-test” (i.e. “How will this affect the implementation of strace,
> when considering a ILP32-compiled and a LP64-compiled strace where both should
> be capable of tracing either ILP32 or LP64 targets?”), option (b) appears the cleanest
> choice, as no test on the dependent’s ABI would be necessary and all internal 
> dispatching could be performed on syscall numbers alone.
> 
> Using the same "strace-test", option (d) would be the least preferred, as it will make 
> the internal dispatching entirely dependent on the dependent’s ABI.

I don't understand what you mean here, please elaborate. Why would an ABI that works
on aarch32 be wrong on aarch64-ilp32 user space when you are using the same header
files?

> >> I was trying to model ilp32 so there was less maintain hassle if a new syscall was added. 
> >> 
> >> Also about time_t, my original patch had used 32bit but was asked to change
> >> it to the 64bit one. So now I am upset this being asked again to change it back.
> >> The review process for the linux kernel is much harder than the review process
> >> of gcc or even glibc now. 
> > 
> > For now, I'm just opening that discussion again, but the reason this
> > comes up again now is that a lot has happened in the meantime on this
> > front, and we have already decided to merge new architecture ports with
> > 32-bit time_t since.
> 
> I think my sloppy e-mail writing blew this out of proportion: I never intended to focus on 
> ‘time_t’, but on ‘timespec’ as a whole (i.e. keeping ‘tv_nsec’ defined as a ‘long’ in 
> userspace). The problem, as far as I can see from the kernel source, is that 
> kernel/compat.c assumes that tv_sec (time_t) and tv_nsec (long) are of equal size.

There are multiple problems relating to time_t here, the padding in
timespec is one of them. There are similar problems relating to __kernel_size_t,
__kernel_ptrdiff_t, __kernel_off_t, __kernel_clock_t, __kernel_ino_t and
structs that build on top of these: whenever you have device driver using
these in an ioctl, you don't know what a user space tool passes, as they
might use a kernel header that contains a 64-bit definition for that type,
or they may have their own definition using the standard types, e.g. copied
from an older kernel, or written independently.

time_t is just special here, because we know that we have to extend it
to 64-bit eventually, while for the others, staying at 32-bit wide types
would generally help compatibility with existing source code bases, both
in kernel drivers and in user space.

> My hope was that to resolve this by extending the compat-layer (in those places where
> COMPAT_USE_64BIT_TIME is tested) with a COMPAT_USE_64BIT_TIME_COMPLIANT
> path that supports a 64-bit time_t with a ILP32-long tv_nsecs.
> 
> We should not break C11-compliant programs by changing the size of tv_nsecs.  As a
> consequence, we shouldn’t propagate a questionable choice made for x32 into ILP32.
> On the other hand, I don’t want to limit time_t to 32bits, as C11 permits any reasonable
> define for it…

Breaking C11 is one concern, but to me the much more valuable question is
whether we break existing user space code. This is about code that is
already not 64-bit clean, so it's also likely to not cope well with
e.g. a 64-bit __kernel_ulong_t.

	Arnd