On 24 June 2013 22:15, Paolo Bonzini <pbonz...@redhat.com> wrote: > Il 24/06/2013 21:21, Ed Maste ha scritto: >> Signed-off-by: Ed Maste <ema...@freebsd.org> >> --- >> I have had this in a local tree for some time, and it is needed by the >> BSD-user work that is now being proposed. > > At this time, qemu/tls.h is really just for cpu_single_env, so I think > this patch should be applied together with the bsd-user patches that > need it.
This patch has arrived because I suggested splitting it out from those ;-) System mode qemu is multi-threaded, so moving more host OSes into the fold of "these variables behave like the linux host which is where the bulk of testing happens" sounds like a good idea to me. >> As an aside, an abstraction was recently proposed for Open vSwtich that >> can use any of _Thread_local, __thread, or pthread_getspecific() which >> may make a convenient reference for someone wishing to implement one of >> the TODOs: http://openvswitch.org/pipermail/dev/2013-June/028665.html > > I and Stefan Hajnoczi have almost the same idea implemented in QEMU > (except that get_foo() returns a pointer to the variable). But > pthread_get/setspecific would be too slow for cpu_single_env, so we're > just switching to __thread for cpu_single_env (for Linux in our patches, > but you can add FreeBSD too once it's needed). I would really like cpu_single_env (and the other thread variables) to have consistent semantics everywhere (which means "always thread local"), even if that means "sometimes suboptimal performance". Reasoning about a variable that might or might not be per-thread is just painful. Alternatively, we could just drop support for any host OS that doesn't provide reasonably efficient per-thread data. -- PMM
