On Fri, 2009-04-10 at 14:08 +0200, Gilles Chanteperdrix wrote:
> Philippe Gerum wrote:
> > On Fri, 2009-04-10 at 11:23 +0200, Gilles Chanteperdrix wrote:
> >> Hi Philippe,
> >>
> >> I got some changes ready for head. What we want to include in the stable
> >> branch remains to be discussed, once we agree, I will prepare another
> >> branch for v2.4.x patches.
> >>
> > 
> > No objection to merge back FPU fixes to 2.4.x before we close that
> > branch, when 2.5 is out. This would give us some time to make sure
> > everything is fine while running the -rc series.
> 
> Ok. I was thinking that it may be dangerous to merge the "Optimize x86
> fpu switches" patch, since it is not strictly necessary. But from the
> point of view of getting it tested ASAP by users, you are right. The
> concern about switchtest is that the changes once again break the driver
> ABI.
> 

I was only referring to the x86-FPU-fixes patch for a merge back to
2.4.x, benefiting from more testing from the 2.5-rc series. The rest
looks like 2.5 business only.

> > 
> >> The following changes since commit 
> >> bbbaec33689d8e82b604745bb55209a83d79a4bc:
> >>   Philippe Gerum (1):
> >>         Test for self-deletion in a safer way
> >>
> >> are available in the git repository at:
> >>
> >>   git://git.xenomai.org/xenomai-gch.git for-upstream
> >>
> >> Gilles Chanteperdrix (4):
> >>       Improve switchtest coverage.
> >>       x86 FPU fixes
> >>       Optimize x86 fpu switches.
> >>       Fix rt_task_trampoline and rt_task_shadow error paths.
> > 
> > I'm generally ok with the patches, but the last one still leaves an
> > issue open: if the child thread dies upon -ENOMEM, the creator won't be
> > unblocked from pending on the completion sync in rt_task_create(). We
> > could live with this for a while (lacking memory at that point is a
> > clear sign that things are going to turn ugly very soon anyway), but
> > would we want to fix this, we would have to either fire the
> > __rt_task_create syscall with some NULL args and let it notice them,
> > then signal the completion block with an error status, or have something
> > like __xn_sys_sigcompletion to unblock the waiter directly from
> > userland.
> 
> I just reused the "fail" label of rt_task_trampoline without thinking
> about the consequences. Will try to see which idea is simpler to
> implement. In any case, I am afraid we will get yet another ABI change.
> 
> A question about git now: can I "git reset" or "git branch -D" a remote
> branch ?
> 
-- 
Philippe.



_______________________________________________
Xenomai-core mailing list
Xenomai-core@gna.org
https://mail.gna.org/listinfo/xenomai-core

Reply via email to