This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: glibc 2.27: less than two weeks till release
On 22/01/18 11:53, Adhemerval Zanella wrote:
> On 22/01/2018 09:11, Szabolcs Nagy wrote:
>> On 21/01/18 22:01, Romain Naour wrote:
>>> 2) aarch64 (gcc):
>>>
>>> libsanitizer/sanitizer_common/sanitizer_linux.cc:1265:35: error: ‘struct
>>> mcontext_t’ has no member named ‘__reserved’; did you mean ‘__glibc_reserved1’?
>>> u8 *aux = ucontext->uc_mcontext.__reserved;
>>> ^~~~~~~~~~
>>> __glibc_reserved1
>>>
>>> Build issue introduced by:
>>> https://sourceware.org/git/?p=glibc.git;a=commit;h=4fa9b3bfe6759c82beb4b043a54a3598ca467289
>>>
>>> * sysdeps/unix/sysv/linux/aarch64/ucontext_i.sym (oEXTENSION): Use
>>> __glibc_reserved1 instead of __reserved.
>>>
>>> Even with the current gcc master, we can't build an aarch64 toolchain with the
>>> upcoming glibc 2.27 since ucontext->uc_mcontext.__reserved is still used.
>>
>> indeed that commit seems faulty:
>>
>> on aarch64 a list of cpu states (e.g fp registers) can only be
>> accessed via parsing the __reserved member (there are no macros
>> or functions defined for this, the user has to manually cast).
>>
>> and __reserved does not violate the namespace rules, so i don't
>> think we need macro hackery to use different name based on
>> _GNU_SOURCE.
>>
>> so i think this api should not be changed, i'll revert this part
>> of that patch.
>>
>> thanks for catching it.
>>
>
> I noted it sometime ago and added a fix on my ILP32 sanitizer support on
> compiler-rt [1] (still in review):
>
> +// glibc starting at 2.26 does not typedef sigcontext to mcontext_t,
> +// but rather defines its own with internal reserved member with
> +// different naming. The mcontext_t definition below is based on
> +// Linux UAPI for sicontex_t.
> +struct __sanitizer_mcontext_t {
> + unsigned long long int fault_address;
> + unsigned long long int regs[31];
> + unsigned long long int pc;
> + unsigned long long int sp;
> + unsigned long long int pstate;
> + unsigned char reserved[4096] __attribute__ ((__aligned__ (16))); // NOLINT
> +};
> +
> static bool Aarch64GetESR(ucontext_t *ucontext, u64 *esr) {
> static const u32 kEsrMagic = 0x45535201;
> - u8 *aux = ucontext->uc_mcontext.__reserved;
> + __sanitizer_mcontext_t *mctx = reinterpret_cast<__sanitizer_mcontext_t *>
> + (&ucontext->uc_mcontext);
> + u8 *aux = mctx->reserved;
casting to struct sigcontext* is fine here (that's the
linux asm/sigcontext.h api), however no existing code
would do that since mcontext_t used to be a typedef of
struct sigcontext.
> while (true) {
> _aarch64_ctx *ctx = (_aarch64_ctx *)aux;
> if (ctx->size == 0) break;
>
>
> I am not sure if we should fix on glibc or sanitizer side.
>
i think glibc should keep the sigcontext and mcontext
fields the same for now (except in posix mode the
namespace violations should be fixed up), there are
likely other users of it than the sanitizers.
>
> [1] https://reviews.llvm.org/D40134
>