Re: [PATCH glibc 1/3] glibc: Perform rseq registration at C startup and thread creation (v19)

[Mathieu Desnoyers]

----- On May 26, 2020, at 10:38 AM, Florian Weimer fweimer@xxxxxxxxxx wrote:

> * Mathieu Desnoyers:
> 
>> AFAIU, the only gain here would be to make sure we don't emit useless
>> ";" in the "/* nothing */" case. But does it matter ?
> 
> I don't think C allows empty constructs like this at the top level.
> 
>>>>> And something similar for _Alignas/attribute aligned,
>>>>
>>>> I don't see where _Alignas is needed here ?
>>>>
>>>> For attribute aligned, what would be the oldest supported C and C++
>>>> standards ?
>>> 
>>> There are no standardized attributes for C, there is only _Alignas.
>>> C++11 has an alignas specifier; it's not an attribute either.  I think
>>> these are syntactically similar.
>>
>> There appears to be an interesting difference between attribute aligned
>> and alignas. It seems like alignas cannot be used on a structure declaration,
>> only on fields, e.g.:
>>
>> struct blah {
>>         int a;
>> } _Alignas (16);
>>
>> o.c:3:1: warning: useless â_Alignasâ in empty declaration
>>  } _Alignas (16);
>>
>> But
>>
>> struct blah {
>>         int _Alignas (16) a;
>> };
> 
> Like the attribute, it needs to come right after the struct keyword, I
> think.  (Trailing attributes can be ambiguous, but not in this case.)

Nope. _Alignas really _is_ special :-(

struct _Alignas (16) blah {
        int a;
};

p.c:1:8: error: expected â{â before â_Alignasâ
 struct _Alignas (16) blah {

Also:

struct blah _Alignas (16) {
        int a;
};

p.c:1:27: error: expected identifier or â(â before â{â token
 struct blah _Alignas (16) {

> 
>> is OK. So if I change e.g. struct rseq_cs to align
>> the first field:
>>
>> struct rseq_cs
>>   {
>>     /* Version of this structure.  */
>>     uint32_t rseq_align (32) version;
>>     /* enum rseq_cs_flags.  */
>>     uint32_t flags;
>>     uint64_t start_ip;
>>     /* Offset from start_ip.  */
>>     uint64_t post_commit_offset;
>>     uint64_t abort_ip;
>>   };
>>
>> It should work.
> 
> Indeed.

OK, so let's go for that approach.

> 
>> /* Rely on GNU extensions for older standards and tls model.  */
>> #ifdef __GNUC__
>> # ifndef rseq_alignof
>> #  define rseq_alignof(x) __alignof__ (x)
>> # endif
>> # ifndef rseq_alignas
>> #  define rseq_alignas(x) __attribute__ ((aligned (x)))
>> # endif
>> # define rseq_tls_model_ie __attribute__ ((__tls_model__ ("initial-exec")))
>> #else
>> /* Specifying the TLS model on the declaration is optional.  */
>> # define rseq_tls_model_ie /* Nothing.  */
>> #endif
>>
>> /* Fall back to __thread for TLS storage class.  */
>> #ifndef rseq_tls_storage_class
>> # define rseq_tls_storage_class __thread
>> #endif
> 
> If they are only used in the glibc headers, they should have __rseq
> prefixes, so that application code doesn't start using them (in case we
> have to change/fix them, or move the into <sys/cdefs.h> later).

OK will do.

> 
> Rest looks fine.

One last thing I'm planning to add in sys/rseq.h to cover acessing the
rseq_cs pointers with both the UAPI headers and the glibc struct rseq
declarations:

/* The rseq_cs_ptr macro can be used to access the pointer to the current
   rseq critical section descriptor.  */
#ifdef __LP64__
# define rseq_cs_ptr(rseq) \
           ((const struct rseq_cs *) (rseq)->rseq_cs.ptr)
#else /* __LP64__ */
# define rseq_cs_ptr(rseq) \
           ((const struct rseq_cs *) (rseq)->rseq_cs.ptr.ptr32)
#endif /* __LP64__ */

Does it make sense ?

Thanks,

Mathieu

> 
> Thanks,
> Florian

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