On Thu, Mar 01, 2018 at 09:10:42AM -0700, Keith Busch wrote: > On Thu, Mar 01, 2018 at 11:03:30PM +0800, Ming Lei wrote: > > If all CPUs for the 1st IRQ vector of admin queue are offline, then I > > guess NVMe can't work any more. > > Yikes, with respect to admin commands, it appears you're right if your > system allows offlining CPU0. > > > So looks it is a good idea to make admin queue's IRQ vector assigned as > > non-managed IRQs. > > It'd still be considered managed even if it's a 'pre_vector', though > it would get the default mask with all possible CPUs.
Which basically does the right thing. So I suspect we'll need to go with a patch like this, just with a way better changelog.