Re: rseq: How to test for compat task at signal delivery

[Mathieu Desnoyers]

----- On Jun 26, 2018, at 3:55 PM, Andy Lutomirski luto@xxxxxxxxxxxxxx wrote:

> On Tue, Jun 26, 2018 at 12:50 PM Mathieu Desnoyers
> <mathieu.desnoyers@xxxxxxxxxxxx> wrote:
>>
>> ----- On Jun 26, 2018, at 3:32 PM, Andy Lutomirski luto@xxxxxxxxxxxxxx wrote:
>>
>> > On Tue, Jun 26, 2018 at 11:45 AM Mathieu Desnoyers
>> > <mathieu.desnoyers@xxxxxxxxxxxx> wrote:
>> >>
>> >> ----- On Jun 26, 2018, at 1:38 PM, Mathieu Desnoyers
>> >> mathieu.desnoyers@xxxxxxxxxxxx wrote:
>> >>
>> >> > Hi Andy,
>> >> >
>> >> > I would like to 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 kernels.
>> >> >
>> >> > I know that internally, the kernel is making a transition from
>> >> > is_compat_task() to in_compat_syscall().
>> >> >
>> >> > I'm fine with using in_compat_syscall() when rseq_get_rseq_cs() is
>> >> > invoked from a system call, but is it OK to call it when it is
>> >> > invoked from signal delivery ? AFAIU, signals can be delivered
>> >> > upon return from interrupt as well.
>> >> >
>> >> > If not, what strategy do you recommend for arch-agnostic code ?
>> >>
>> >> I think what we're missing here is a new "is_compat_frame(struct ksignal *ksig)"
>> >> which I could use in the rseq code. I'll prepare a patch and we can discuss
>> >> from there.
>> >>
>> >
>> > That sounds about right.
>> >
>> > I'm confused, though.  Wouldn't it be more consistent to just segfault
>> > if the high 32 bits are not clear when rseq transitions to a 32-bit
>> > context?  If there's garbage in 64-bit mode, the program will crash.
>> > Why should 32-bit mode be any different?
>>
>> Currently, if a 32-bit binary puts garbage in the high bits of
>> start_ip, post_commit_offset, and abort_ip in
>>
>> include/uapi/linux/rseq.h:
>>
>> struct rseq_cs {
>>         /* Version of this structure. */
>>         __u32 version;
>>         /* enum rseq_cs_flags */
>>         __u32 flags;
>>         LINUX_FIELD_u32_u64(start_ip);
>>         /* Offset from start_ip. */
>>         LINUX_FIELD_u32_u64(post_commit_offset);
>>         LINUX_FIELD_u32_u64(abort_ip);
>> } __attribute__((aligned(4 * sizeof(__u64))));
> 
> This ABI isn't real ABI until a stable kernel happens, right?  So how
> about just making all those fields be u64?

Good point. Unlike the rseq_cs field in the struct rseq TLS, those
fields don't need to be word-sized/word-aligned, so we could simply
declare them as __u64.

> 
>>
>> A 32-bit kernel just never reads the padding, thus in reality acting
>> as if those were zeroes. However, a 64-bit kernel dealing with this
>> 32-bit compat task will read that padding, handling those as very
>> large values.
> 
> Sounds like a design error.  Have all kernels read the fields no
> matter what.  A 32-bit kernel will send SIGSEGV if the high bits are
> set.  A 64-bit kernel running compat userspace should make sure that a
> 32-bit task dies if the high bits are set.

If we end up declaring those as __u64, that approach makes sense.

> 
>>
>> We need to improve that by introducing a consistent behavior across
>> native 32-bit kernels and 32-bit compat mode on 64-bit kernels.
>>
>> There are two ways to achieve this: either the 32-bit kernel validates
>> the padding by killing the process if padding is non-zero, or the
>> 64-bit kernel treats compat mode by zeroing the high bits of padding.
>>
>> If we look at system call interfaces in general, I think the usual
>> approach is to clear the top bits whenever a value read from a
>> compat task ends up being used as a pointer. This is why I am tempted
>> to go for the "clear high bits" approach rather than killing the task.
> 
> I think the modern preference is to use fields of fixed size rather
> than long when UABI is involved.
> 
> In any event, I think the test you want is user_64bit_mode().

Currently, user_64bit_mode is only implemented on x86.

Should we introduce an architecture-agnostic user_64bit_mode(struct pt_regs *)
which maps to is_compat_task() for non-x86 ? I'm just worried that ptrace
code could try to use it from the context of another task and get mixed up.

> 
>>
>> Also, validating that the top 32-bit is zeroes from a native 32-bit
>> kernel requires extra loads, whereas not caring about their content
>> is free, which makes me slightly prefer an approach where 32-bit
>> compat mode on 64-bit kernel just clears the top bits.
>>
> 
> But performance is totally irrelvant here, right?  This only affects
> the abort path, unless I'm rather confused.

This load is added on return-to-userspace after a preemption, and upon
signal delivery. So it's not a fast-path.

Thanks,

Mathieu


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