On Thu, 24 Mar 2016 09:47:56 +0100 Cornelia Huck <cornelia.h...@de.ibm.com> wrote:
> On Thu, 24 Mar 2016 09:32:30 +0100 > Cornelia Huck <cornelia.h...@de.ibm.com> wrote: > > > I'll re-send my patches with some added interface doc in patch 1. Stay > > tuned. > > Grr. Unfortunately, this fails for _me_ now (-EEXIST after reboot). > Debugging. Good news is that I think I understand what happens. Bad news is that we can scratch all of the previous testing :( My patchset had a typo (check for !disabled instead of disabled). After I fixed that, the second assignment of the ioeventfd started failing (that's what changed when I started passing assign in stop_ioeventfd) with -EEXIST as the previous ioeventfd is of course still assigned. What we actually want is to keep the ioeventfd assigned, not add a new one. But we still want adding a new ioeventfd to fail in case of collisions. I think we need to track whether we already assigned an ioeventfd and don't re-register in that case. I'll try to cook something up.