Date:   Fri, 20 Mar 2020 09:44:07 -0400 (EDT)
From:   Mathieu Desnoyers <mathieu.desnoyers@...icios.com>
To:     Florian Weimer <fw@...eb.enyo.de>
Cc:     libc-alpha <libc-alpha@...rceware.org>, carlos <carlos@...hat.com>,
        Rich Felker <dalias@...c.org>,
        linux-api <linux-api@...r.kernel.org>,
        Boqun Feng <boqun.feng@...il.com>,
        Will Deacon <will.deacon@....com>,
        linux-kernel <linux-kernel@...r.kernel.org>,
        Peter Zijlstra <peterz@...radead.org>,
        Ben Maurer <bmaurer@...com>, Dave Watson <davejwatson@...com>,
        Thomas Gleixner <tglx@...utronix.de>,
        Paul <paulmck@...ux.vnet.ibm.com>, Paul Turner <pjt@...gle.com>,
        Joseph Myers <joseph@...esourcery.com>
Subject: Re: [RFC PATCH glibc 4/8] glibc: Perform rseq(2) registration at C
 startup and thread creation (v15)

----- On Mar 19, 2020, at 3:46 PM, Mathieu Desnoyers mathieu.desnoyers@...icios.com wrote:

> ----- On Mar 19, 2020, at 3:05 PM, Florian Weimer fw@...eb.enyo.de wrote:
> 
>> * Mathieu Desnoyers:
>> 
>>>> Inside glibc, you can assume __attribute__ support.
>>>
>>> OK, so the _Static_assert () could sit in sys/rseq.h
>> 
>> It requires a C11 compiler.  In this case, you could use _Alignas.
> 
> How would _Alignas replace:
> 
> +_Static_assert (__alignof__ (struct rseq_cs) >= 4 * sizeof(uint64_t),
> +                "alignment");
> +_Static_assert (__alignof__ (struct rseq) >= 4 * sizeof(uint64_t),
> +                "alignment");
> 
> ?
> 
> Moreover, I notice that sys/cdefs.h implements a fallback for _Static_assert
> for cases where it is not supported by the compiler. So I do not think it
> strictly depends on C11 if I include sys/cdefs.h from sys/rseq.h.
> 
>>>>>>>> The struct rseq/struct rseq_cs definitions
>>>>>>>> are broken, they should not try to change the alignment.
>>>>>>>
>>>>>>> AFAIU, this means we should ideally not have used __attribute__((aligned))
>>>>>>> in the uapi headers in the first place. Why is it broken ?
>>>>>> 
>>>>>> Compilers which are not sufficiently GCC-compatible define
>>>>>> __attribute__(X) as the empty expansion, so you silently get a
>>>>>> different ABI.
>>>>>
> 
> [...]
> 
>>>>>> There is really no need to specify 32-byte alignment here.  Is not
>>>>>> even the size of a standard cache line.  It can result in crashes if
>>>>>> these structs are heap-allocated using malloc, when optimizing for
>>>>>> AVX2.
>>>>>
>>>>> Why would it be valid to allocate those with malloc ? Isn't it the
>>>>> purpose of posix_memalign() ?
>>>> 
>>>> It would not be valid, but I don't think we have diagnostics for C
>>>> like we have them for C++'s operator new.
>>>
>>> We could at least make an effort to let people know that alignment is
>>> required here when allocating struct rseq and struct rseq_cs on the
>>> heap by adding some comments to that effect in linux/rseq.h ?
>> 
>> We could use different types on the glibc side, then no special
>> programmer action will be needed.
> 
> Can't this lead to problems when mixing up compile units which have
> been compiled with linux/rseq.h with compile units compiled against
> sys/rseq.h ?
> 
> Let me take a step back and try to understand.
> 
> So far, there appears to be two scenarios where having a 64-byte
> alignment attribute on struct rseq and struct rseq_cs can cause
> problems:
> 
> 1) A user-space programmer uses malloc() to dynamically allocate
>   struct rseq or struct rseq_cs, which does not satisfy any of
>   the alignment requirement of the structure. Combining this with
>   compiler expectations that the structure needs to be aligned
>   on 64-byte (e.g. -mavx2) breaks things.
> 
>   For this first scenario, I am proposing that we document that
>   the programmer should have used posix_memalign(), which provides
>   the required alignment guarantees.
> 
> 2) A user-space programmer mixes code compiled with compilers
>   honouring the aligned attribute with other compile units compiled
>   with compilers which discard those GCC extension attributes silently,
>   embeds those into a structure, and get different struct layouts.
> 
>   The _Static_assert in sys/rseq.h should detect the case where a
>   compiler is not honouring the aligned attribute, right ?
> 
>> 
>>>>>>> However, now that it is in the wild, it's a bit late to change that.
>>>>>> 
>>>>>> I had forgotten about the alignment crashes.  I think we should
>>>>>> seriously consider changing the types. 8-(
>>>>>
>>>>> I don't think this is an option at this stage given that it is part
>>>>> of the Linux kernel UAPI. I am not convinced that it is valid at all
>>>>> to allocate struct rseq or struct rseq_cs with malloc(), because it
>>>>> does not guarantee any alignment.
>>>> 
>>>> The kernel ABI doesn't change.  The kernel cannot use the alignment
>>>> information anyway.  Userspace struct layout may change in subtle
>>>> ways, though.
>>>
>>> Considering the amount of pain this can cause in user-space, and because
>>> it can break userspace, this is not a UAPI change I am willing to consider.
>>> I'm not sure why we are even discussing the possibility of breaking a Linux
>>> UAPI considering that those are set in stone.
>> 
>> Again, the kernel interface is NOT affected.  Only if the struct is
>> used in a non-top-level fashion across an ABI boundary in userspace.
>> I think making the change now is better than dealing with the breakage
>> in rseq users when they are built with -mavx2.
> 
> What I am missing is what are the issues that persist once we add proper
> documentation of alignment requirements for heap allocation and a static
> assert to fail early when compiled with a compiler dismissing the
> aligned attribute ?
> 
> As you point out, changing the currently public linux/rseq.h UAPI header
> to remove those attributes ends up breaking user-space in scenarios of
> non-top-level use across ABI boundary. This is not kernel-vs-userspace
> ABI, but an ABI exposed by the kernel which ends up being used to
> coordinate user-space objects within a program. Breaking that does not
> appear to be any more acceptable. As I recall, the hard requirement for
> Linux ABIs is to do not break userspace, period. There is not mention
> of kernel-vs-userspace or userspace-vs-userspace. So if the end result
> of this change is to break user-space, it should not be changed.

Actually, here is an important clarification: the Linux kernel validates
the struct rseq alignment on registration:

        if (!IS_ALIGNED((unsigned long)rseq, __alignof__(*rseq)) ||
            rseq_len != sizeof(*rseq))
                return -EINVAL;

So removing the aligned attribute from struct rseq is actually an
ABI-breaking change, because it would be incompatible with older
kernels which perform the IS_ALIGNED check expecting at least at
32 bytes alignment.

Thanks,

Mathieu

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

Powered by blists - more mailing lists