Message-ID: <858886246.10882.1530583291379.JavaMail.zimbra@efficios.com>
Date:   Mon, 2 Jul 2018 22:01:31 -0400 (EDT)
From:   Mathieu Desnoyers <mathieu.desnoyers@...icios.com>
To:     Andy Lutomirski <luto@...capital.net>
Cc:     Linus Torvalds <torvalds@...ux-foundation.org>,
        Thomas Gleixner <tglx@...utronix.de>,
        linux-kernel <linux-kernel@...r.kernel.org>,
        linux-api <linux-api@...r.kernel.org>,
        Peter Zijlstra <peterz@...radead.org>,
        "Paul E. McKenney" <paulmck@...ux.vnet.ibm.com>,
        Boqun Feng <boqun.feng@...il.com>,
        Dave Watson <davejwatson@...com>, Paul Turner <pjt@...gle.com>,
        Andrew Morton <akpm@...ux-foundation.org>,
        Russell King <linux@....linux.org.uk>,
        Ingo Molnar <mingo@...hat.com>,
        "H. Peter Anvin" <hpa@...or.com>, Andi Kleen <andi@...stfloor.org>,
        Chris Lameter <cl@...ux.com>, Ben Maurer <bmaurer@...com>,
        rostedt <rostedt@...dmis.org>,
        Josh Triplett <josh@...htriplett.org>,
        Catalin Marinas <catalin.marinas@....com>,
        Will Deacon <will.deacon@....com>,
        Michael Kerrisk <mtk.manpages@...il.com>,
        Joel Fernandes <joelaf@...gle.com>
Subject: Re: [RFC PATCH for 4.18] rseq: use __u64 for rseq_cs fields,
 validate user inputs

----- On Jul 2, 2018, at 9:19 PM, Mathieu Desnoyers mathieu.desnoyers@...icios.com wrote:

> ----- On Jul 2, 2018, at 7:37 PM, Andy Lutomirski luto@...capital.net wrote:
> 
>>> On Jul 2, 2018, at 4:22 PM, Mathieu Desnoyers <mathieu.desnoyers@...icios.com>
>>> wrote:
>>> 
>>> ----- On Jul 2, 2018, at 7:16 PM, Mathieu Desnoyers
>>> mathieu.desnoyers@...icios.com wrote:
>>> 
>>>> ----- On Jul 2, 2018, at 7:06 PM, Linus Torvalds torvalds@...ux-foundation.org
>>>> wrote:
>>>> 
>>>>> On Mon, Jul 2, 2018 at 4:00 PM Mathieu Desnoyers
>>>>> <mathieu.desnoyers@...icios.com> wrote:
>>>>>> 
>>>>>> Unfortunately, that rseq->rseq_cs field needs to be updated by user-space
>>>>>> with single-copy atomicity. Therefore, we want 32-bit user-space to initialize
>>>>>> the padding with 0, and only update the low bits with single-copy atomicity.
>>>>> 
>>>>> Well... It's actually still single-copy atomicity as a 64-bit value.
>>>>> 
>>>>> Why? Because it doesn't matter how you write the upper bits. You'll be
>>>>> writing the same value to them (zero) anyway.
>>>>> 
>>>>> So who cares if the write ends up being two instructions, because the
>>>>> write to the upper bits doesn't actually *do* anything.
>>>>> 
>>>>> Hmm?
>>>> 
>>>> Are there any kind of guarantees that a __u64 update on a 32-bit architecture
>>>> won't be torn into something daft like byte-per-byte stores when performed
>>>> from C code ?
>>>> 
>>>> I don't worry whether the upper bits get updated or how, but I really care
>>>> about not having store tearing of the low bits update.
>>> 
>>> For the records, most updates of those low bits are done in assembly
>>> from critical sections, for which we control exactly how the update is
>>> performed.
>>> 
>>> However, there is one helper function in user-space that updates that value
>>> from C through a volatile store, e.g.:
>>> 
>>> static inline void rseq_prepare_unload(void)
>>> {
>>>        __rseq_abi.rseq_cs = 0;
>>> }
>> 
>> How about making the field be:
>> 
>> union {
>> __u64 rseq_cs;
>> struct {
>>   __u32 rseq_cs_low;
>>   __u32 rseq_cs_high;
>> };
>> };
>> 
>> 32-bit user code that cares about performance can just write to rseq_cs_low
>> because it already knows that rseq_cs_high == 0.
>> 
>> The header could even supply a static inline helper write_rseq_cs() that
>> atomically writes a pointer and just does the right thing for 64-bit, for
>> 32-bit BE, and for 32-bit LE.
>> 
>> I think the union really is needed because we can’t rely on user code being
>> built with -fno-strict-aliasing.  Or the helper could use inline asm.
>> 
>> Anyway, the point is that we get optimal code generation (a single instruction
>> write of the correct number of bits) without any compat magic in the kernel.
> 
> That works for me! Any objection from anyone else for this approach ?

One thing to consider is how we will implement the load of that pointer
on the kernel side. Strictly-speaking, the rseq uapi talks about single-copy
atomicity, and does not specify _which_ thread is expected to update that
pointer. So arguably, the common case is that the current thread is updating
it, which would allow the kernel to read it piece-wise. However, nothing
prevents user-space from updating it from another thread with single-copy
atomicity.

So in order to be on the safe side, I prefer to guarantee single-copy
atomicity of the get_user() load from the kernel that reads this pointer.
This means a 32-bit kernel would have to perform two independent loads:
one for low bits, one for high bits.

So it does look like we need some __LP64__ ifdefery even with the union
trick. Therefore, I'm not convinced the union is useful at all.

Thoughts ?

Thanks,

Mathieu

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

Powered by blists - more mailing lists