Re: [Qemu-devel] [PATCH] Make cpu_single_env thread local (Linux only for now)

[Jan Kiszka <jan.kiszka@siemens.com>]

On 2011-10-26 17:09, Peter Maydell wrote:
> On 26 October 2011 16:02, Jan Kiszka <jan.kiszka@siemens.com> wrote:
>> On 2011-10-26 16:54, Peter Maydell wrote:
>>> On 26 October 2011 15:39, Paolo Bonzini <pbonzini@redhat.com> wrote:
>>>> On 10/26/2011 04:03 PM, Peter Maydell wrote:
>>>>>
>>>>> For the record (since I think we only talked about this on IRC):
>>>>>  * the POSIX TLS fallback code doesn't work on Linux hosts for
>>>>>    linux-user emulation (the constructor is never called to set up
>>>>>    the TLS for the main thread, probably something to do with our
>>>>>    custom linker script, since it does work OK for system emulation)
>>>>>  * if I recall correctly from IRC it doesn't compile on OpenBSD
>>>>>
>>>>> so for 1.0 perhaps we need to fall back to a simpler set of patches
>>>>> that just avoid the regression in thread support for linux-user ?
>>>>
>>>> I agree.
>>>
>>> I was thinking something like a trivial qemu-tls.h which does
>>>
>>> #ifdef __linux__
>>> #define DECLARE_TLS(type, x) extern DEFINE_TLS(type, x)
>>> #define DEFINE_TLS(type, x)  __thread __typeof__(type) tls__##x
>>> #define get_tls(x)           tls__##x
>>> #else
>>> /* Dummy implementations -- we can get away with this because system
>>>  * mode is effectively single-threaded for our current limited use of
>>>  * TLS, and the only -user mode which supports multiple threads is
>>>  * linux-user.
>>
>> We will use it in system mode as well, e.g. when I'll push away the
>> global lock from KVM entry/exits. So let's state that it is only
>> interesting on Linux hosts for now.
>>
>> Otherwise, this looks good.
> 
> The point of the comment is that you can't use this facility in
> system mode until we've added the non-Linux support (because you'd
> break Win32 &c). So fixing and committing the non-Linux support is
> a prerequisite for anything like the global lock moves. I can be
> a bit more explicit about that if you like:
> 
> /* Dummy implementations -- we can get away with this because system
>  * mode is effectively single-threaded for our current limited use of
>  * TLS, and the only -user mode which supports multiple threads is
>  * linux-user.
>  * This means you cannot use this for any variables which will
>  * actually be accessed by more than one thread in system mode
>  * until the implementations for Win32 and POSIX systems without
>  * __thread have been added!
>  *
>  * TODO: add implementations via Win32 .tls sections and
>  * POSIX pthread_getspecific.
>  */
> 
> -- PMM

My point is that it is fine to use for per-vcpu variables because:
 - they are single-threaded in TCG mode
 - they are multi-threaded in KVM mode, but that's only affecting Linux
   hosts for which this TLS wrapper is already usable

Jan

-- 
Siemens AG, Corporate Technology, CT T DE IT 1
Corporate Competence Center Embedded Linux