On Sat, Oct 13, 2018 at 10:34:11AM +0100, Catalin Marinas wrote:
>
> Lines: 73
>
> External Email
>
> On Sat, Oct 13, 2018 at 04:14:16AM +0200, Eugene Syromiatnikov wrote:
> > On Wed, Oct 10, 2018 at 04:36:56PM +0100, Catalin Marinas wrote:
> > > On Wed, Oct 10, 2018 at 04:10:21PM +0200, Eugene Syromiatnikov wrote:
> > > > I have some questions regarding AArch64 ILP32 implementation for which I
> > > > failed to find an answer myself:
> > > > * How ptrace() tracer is supposed to distinguish between ILP32 and LP64
> > > > tracees? For MIPS N32 and x32 this is possible based on syscall
> > > > number, but for AArch64 ILP32 I do not see such a sign. There's also
> > > > ARM_ip is employed for signalling entering/exiting, I wonder whether
> > > > it's possible to employ it also for signalling tracee's personality.
> > >
> > > With the current implementation, I don't think you can distinguish. From
> > > the kernel perspective, the register set is the same. What is the
> > > use-case for this?
> >
> > Err, a ptrace()-based tracer trying to trace a process, for example?
>
> I first thought it wouldn't matter for ptrace-based tracers since the
> syscall numbers are (mostly) the same. But the arguments layout in
> register is indeed different, so I see your point now about having to
> distinguish.
>
> > > We could add a new regset to expose the ILP32 state (NT_ARM_..., I can't
> > > think of a name now but probably not PER* as this implies PER_LINUX_...
> > > which is independent from TIF_32BIT_*).
> >
> > So that would require an additional ptrace() call on each syscall stop,
> > is that correct?
>
> The ILP32 state does not change at run-time, so it could only do a
> ptrace() call once and save the information. No need to re-read it on
> each syscall stop.
>
> We could set a high bit in the syscall number reported to the ptrace
> caller (though not changing the syscall ABI) but I haven't thought of
> other consequences. For example, can the ptrace caller change the
> syscall number?
I believe, /proc/PID/auxv is enough to distinguish between arm64, ilp32
and aarch32 ABis. If no, I think it's better to do it there.
I don't have ILP32 machine available at the moment, but I'll check it soon.
> > > > * What's the reasoning behind capping syscall arguments to 32 bit? x32
> > > > and MIPS N32 do not have such a restriction (and do not need special
> > > > wrappers for syscalls that pass 64-bit values as a result, except
> > > > when they do, as it is the case for preadv2 on x32); moreover, that
> > > > would lead to insurmountable difficulties for AArch64 ILP32 tracers
> > > > that try to trace LP64 tracees, as it would be impossible to pass
> > > > 64-bit addresses to process_vm_{read,write} or ptrace PEEK/POKE.
> > >
> > > We've attempted in earlier versions to allow a mix of 32 and 64-bit
> > > register values from ILP32 but it got pretty complicated. The entry code
> > > would need to know which registers need zeroing of the top 32-bit
> >
> > If kernel specifies 64-bit wide registers for syscalls, then it's the
> > caller's (libc's) responsibility to properly sign-extend arguments when
> > needed, and glibc, for example, already has proper type definitions that
> > aimed to handle this.
>
> We tried, see my other reply.
A couple of links to recall the story:
https://www.spinics.net/lists/linux-s390/msg11593.html
http://linux-kernel.2935.n7.nabble.com/RFC6-PATCH-v6-00-21-ILP32-for-ARM64-td1345105.html
Cover-letter of the series has links to previous discussions.
I would also notice that even if we pass 64-bit parameters in a single
register, we cannot avoid using the compat layer. It looks more natural
not to split the 64-bit register, but from performance point of view
there is almost no difference, either we split registers or not (2.6%
for empty syscall, as I measured). And the cost of overcomplication was
considered too much. So we chose to stick to more standard compat layer
and gain in maintainability.
Yury
