Re: [PATCH 0/3 v2] arm64/ptrace: allow to get all registers on syscall traps

[Keno Fischer <keno@juliacomputing.com>]

>
> Could you describe the problem in more details? I wonder whether we have
> the same thing in CRIU...
>

Sure, basically the issue is that orig_x0 gets recorded at the start of the
syscall entry, but is not otherwise part of the ptrace state. It used
to be primarily used for resetting the argument back to its original
value during syscall restarts, but I see it's been expanded slightly
with the user dispatch mechanism (though as far as I can tell
not in a way that interacts with ptrace). Basically the problem
is that if you change the value of x0 during either the ptrace
entry stop or a seccomp stop and then incur a syscall restart,
the syscall will restart with the original x0 rather than with
the modified x0, which may be unexpected. Of course,
relatedly, if you're doing CRIU-like things you can end up
in situations where the future behavior will depend on the
orig_x0 value, which isn't restore-able at the moment. It's
possible to work around all of this by keeping a local copy
of orig_x0 and being very careful with the ptrace traps around
restarts, but getting the logic right is extremely tricky. My
suggestion for what I thought would be reasonable
behavior was:

1. Expose orig_x0 to ptrace
2. Set orig_x0 to x0 and set x0 to -ENOSYS at the start of the syscall
dispatcher
3. Use orig_x0 for syscall arguments/seccomp/restarts

That's basically how rax works on x86_64 and it doesn't
seem to cause major problems (though of course I may
be biased by having x86_64 already work when I started the
aarch64 port). Just the first item would be sufficient of course
for getting rid of most of the bookkeeping. I should also say
that, for us, the ptrace getregs call can be the throughput
limiting operation, so it would be nice if getting the entire
basic register set would only require one syscall. I won't
insist on it, since we do have a solution in place that kinda
works (and only requires the one syscall),
but I thought I'd mention it.

While we're on this topic, and in case it's helpful to anybody,
I should also point out that the order of the ptrace-signal-stop,
vs setup for the syscall restart differs between x86 and
aarch64 (aarch64 sets up the restart first then delivers the
ptrace trap/signal - x86 the other way around). I actually
think the aarch64 behavior is saner here, but I figured I'd
leave this breadcrumb for anybody who's writing a ptracer
and stumbles across this.

Keno