Re: [RFC PATCH for 4.18 2/2] rseq: compat: clear high bits of rseq_cs fields

[Mathieu Desnoyers]

----- On Jun 26, 2018, at 5:58 PM, Andy Lutomirski luto@xxxxxxxxxxxxxx wrote:

>> On Jun 26, 2018, at 2:16 PM, Mathieu Desnoyers <mathieu.desnoyers@xxxxxxxxxxxx>
>> wrote:
>> 
>> Make the behavior rseq on compat tasks more robust by ensuring that
>> kernel/rseq.c:rseq_get_rseq_cs() clears the high bits of
>> rseq_cs->abort_ip, rseq_cs->start_ip and rseq_cs->post_commit_offset
>> when a 32-bit binary is run on a 64-bit kernel.
>> 
>> The intent here is that if user-space has garbage rather than zeroes
>> in its struct rseq_cs fields padding, the behavior will be the same
>> whether the binary is run on 32-bit or 64-bit kernels.
>> 
>> Use in_compat_syscall() when rseq_get_rseq_cs() is invoked from
>> system call context, and use is_compat_frame() when invoked from
>> signal delivery.
>> 
> 
> And when itâs invoked due to preemption unrelated to a syscall or signal, you
> malfunction?

Fair point! Hence the "RFC". ;)

So I understand better your intent to use the pt_regs to figure out whether it
is compat or not. My is_compat_frame()+in_compat_syscall() approach does not
handle this correctly.

> 
> I think the only sane solution is to make these fields be u64,

I'm OK with turning the rseq_cs start_ip, post_commit_offset, and abort_ip
fields into normal u64.

> delete the
> LINUX_FIELD_ macros,

The LINUX_FIELD_ macros are still needed to ensure single-copy updates of
the (struct rseq *__tls_abi)->rseq_cs pointer by 32-bit user-space.

> and possibly teach the x86 slowpath return to inject a
> signal if itâs trying to return to a 32-bit context with garbage in the high
> bits of regs->ip so that we determistically fail if the user screws up.

I like the approach of dealing with the rseq_cs fields as u64 even on 32-bit
architectures. As a downside, it will require 32-bit architectures to do
arithmetic on 64-bit values, but it's not a fast-path. As you point out, the
tricky bit is to decide what happens when architecture code returns to
userspace with regs->ip containing garbage in the high bits.

An alternative approach is to ensure the high bits are cleared when returning
to an IP with garbage in the high bits.

> Rseq is brand new. It should not need compat code at all.

Dealing with u64 for start_ip, post_commit_offset, and abort_ip at the kernel
level would indeed provide this characteristic. However, I'm uneasy adding
64-bit arithmetic on operations really caring about 32 bits on 32-bit archs,
even though those are not fast paths.

Thanks,

Mathieu

-- 
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com