On Thu, Nov 4, 2021 at 10:20 AM Christoph Hellwig <[email protected]> wrote: > > On Thu, Nov 04, 2021 at 10:17:31AM -0700, Christoph Hellwig wrote: > > On Thu, Nov 04, 2021 at 10:09:37AM -0700, Dan Williams wrote: > > > Hmm, so the driver has 2 modes "labeled" and "label-less", in the > > > labeled mode it waits for an explicit: > > > > > > ndctl create-namespace > > > > > > ...to provision region capacity into a namespace. In label-less mode > > > it just assumes that the boundaries of the region are the boundaries > > > of the namespace. In this case it looks like the driver found a label > > > index block with no namespaces defined so it's waiting for one to be > > > created. Are you saying that the only thing you changed from a working > > > config with defined namespace to this one was a kernel change? I.e. > > > the content of those memory-backend files has not changed? > > > > Well, the config change is the only thing I though of as relevant. > > The content of those files actually changes all the time, as I also > > use them as the backing store for my qemu configs that use block > > devices. E.g. the previous run they did show up as NVMe devices. > > Ok, blowing the files away and recreating them makes the pmem > devices show up again. >
Cool, yeah, I was going to guess the label area might have been clobbered, but there was just enough left to force the driver into labeled mode. You can verify with something like: ndctl read-labels -I -j all ...to see if nvdimms have label index blocks.
