On Tue, Sep 14, 2021 at 09:32:36AM +0200, Christoph Hellwig wrote:

On Tue, Sep 14, 2021 at 10:09:43AM +0300, Sagi Grimberg wrote:

quoted

quoted

quoted

I agree it should not happen with xarray as inserts preserve sorting,
but that patch has the non-trivial removal of the namespaces_rwsem which
synchronizes in a lot in a lot of error recovery flows. It's very
possible that xarray will introduce regressions.

I've actually started looking over the series earlier today, and staring
it I'm pretty sure the namespaces_rwsem is completely broken.  There
is nothing preventing the namespaces from going away in these loops.

Where? the ns itself? the ns is removed from ctrl->namespaces list under
this rwsem (nvme_ns_remove).

Sorry, with the series I meant the xarray conversion that removes the
rwsem entirely.

quoted

quoted

But I can't see why we don't just always do a sorted insertation into
the ->namespaces list.  Just trying to figure out the history of the
delayed sort ATM.

Don't know the history.

git-blame teels me it showed up with the initial ID ns list support
from Keith, so not all that much history here.

The main reason I suggest applying this as a short term fix
is that it is well isolated and is easy to cherry-pick into the LTS/stable versions.

_______________________________________________
Linux-nvme mailing list
Linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme