On Sat, 2015-02-07 at 21:00 +0100, Paolo Bonzini wrote: > > On 07/02/2015 02:39, Alex Williamson wrote: > > I'm not sure where it's coming from yet, but I did extensive testing for > > my last pull request based on ec6f25e because if I updated to d5fbb4c > > vfio hotplug broke immediately. I'll keep looking, but I thought I'd > > share in case you have some ideas. Thanks, > > I'm not sure I understand: d5fbb4c9ed52d97aebe5994d8a857c74c0d95a92 (RCU > merge) is an ancestor of ec6f25e788ef57ce1e9f734984ef8885172fd9e2 (s390 > merge) and the only patches in the middle are for s390.
Ok, I went back to 83761b9244ad, applied 3a4dbe6aa934 to get the object_unparent() fix, then applied this series. Everything seems to work ok. Then I manually applied and bisected the commits that came in via d5fbb4c9ed52. I land on 374f2981d1f1 as introducing the segfault in memory_listener_register(). I guess I was mis-remembering where I did my testing for the last vfio pull request. My tag was based on ec6f25e, but I remember that I had to test based on a commit before the RCU merge. My test is to simply do virsh detach-device, attach-device in a loop for a vfio assigned VF NIC, 1s delay between ops. It typically fails within 100 to 150 iterations, I call 500 a pass. Thanks, Alex