Re: [Qemu-devel] [PATCH 5/5] linux-user/signal.c: define __SIGRTMIN/MAX for non-GNU platforms

[Eric Blake]

On 04/29/2014 08:53 AM, Natanael Copa wrote:
> On Tue, 29 Apr 2014 08:28:29 -0600
> Eric Blake <address@hidden> wrote:
> 
>> On 04/29/2014 08:17 AM, Natanael Copa wrote:
>>> The __SIGRTMIN and __SIGRTMAX are glibc internals and are not available
>>> on all platforms, so we define those if they are missing.

>>> +#define __SIGRTMIN 32
>>
>> Rather than defining the implementation-specific __SIGRTMIN to a magic
>> number that is liable to be wrong, why not instead fix the code to use
>> the POSIX-mandated SIGRTMIN and SIGRTMAX public defines instead?
>>
> 
> Those seems to be runtime values:
> /usr/include/signal.h:#define SIGRTMIN  (__libc_current_sigrtmin())

Oh right - POSIX allows them to be runtime variable.  But we are
interacting with a given kernel, where the values will be fixed.  Maybe
you have to define __SIGRTMIN after all, but can we at least have an
assert() that the value you picked matches SIGRTMIN at runtime?

> /usr/include/signal.h:#define SIGRTMAX  (__libc_current_sigrtmax())
> 
> so it gives:
> /home/ncopa/src/qemu/linux-user/signal.c:93:5: error: nonconstant array index 
> in initializer
>      [SIGRTMIN] = __SIGRTMAX,
> 
> I could have used (NSIG-1) but are not sure if NSIG is a runtime macro
> in glibc. The array itself is using _NSIG instead of NSIG for some
> reason.

NSIG is not any more portable; nor does POSIX require that the RT
signals occur at the tail end of NSIG (in other words, NSIG-1 need not
be SIGRTMAX).  Unless someone knows of a kernel define, it sounds like
we're stuck hard-coding in some knowledge of Linux.

-- 
Eric Blake   eblake redhat com    +1-919-301-3266
Libvirt virtualization library http://libvirt.org

signature.asc
Description: OpenPGP digital signature