Messages in this thread | | | Date | Tue, 9 Nov 2021 10:43:58 -0400 | From | Jason Gunthorpe <> | Subject | Re: [PATCH net-next] devlink: Require devlink lock during device reload |
| |
On Mon, Nov 08, 2021 at 03:31:26PM -0800, Jakub Kicinski wrote: > On Mon, 8 Nov 2021 21:58:36 +0200 Leon Romanovsky wrote: > > > > > nfp will benefit from the simplified locking as well, and so will bnxt, > > > > > although I'm not sure the maintainers will opt for using devlink framework > > > > > due to the downstream requirements. > > > > > > > > Exactly why devlink should be fixed first. > > > > > > If by "fixed first" you mean it needs 5 locks to be added and to remove > > > any guarantees on sub-object lifetime then no thanks. > > > > How do you plan to fix pernet_ops_rwsem lock? By exposing devlink state > > to the drivers? By providing unlocked version of unregister_netdevice_notifier? > > > > This simple scenario has deadlocks: > > sudo ip netns add n1 > > sudo devlink dev reload pci/0000:00:09.0 netns n1 > > sudo ip netns del n1 > > Okay - I'm not sure why you're asking me this. This is not related to > devlink locking as far as I can tell. Neither are you fixing this > problem in your own RFC. > > You'd need to tell me more about what the notifier is used for (I see > RoCE in the call trace). I don't understand why you need to re-register > a global (i.e. not per netns) notifier when devlink is switching name > spaces.
This becomes all entangled in the aux device stuff we did before.
devlink reload is defined, for reasons unrelated to netns, to do a complete restart of the aux devices below the devlink. This happens necessarily during actual reconfiguration operations, for instance.
So we have a situation, which seems like bad design, where reload is also triggered by net namespace change that has nothing to do with reconfiguring. In this case the per-net-ns becomes a BKL that gets held across way too much stuff as it recuses down the reload path, through aux devices, into the driver core and beyond.
When I looked at trying to fix this from the RDMA side I could not find any remedy that didn't involve some kind of change in netdev land. The drivers must be able to register/unregister notifiers in their struct device_driver probe/remove functions.
I once sketched out fixing this by removing the need to hold the per_net_rwsem just for list iteration, which in turn avoids holding it over the devlink reload paths. It seemed like a reasonable step toward finer grained locking.
Jason
| |