On Thu, Jun 13, 2024 at 07:13:43AM -0700, Jakub Kicinski wrote: > On Thu, 13 Jun 2024 10:54:12 +0200 Larysa Zaremba wrote: > > > > The locking mechanisms I use here do not look pretty, but if I am not > > > > missing > > > > anything, the synchronization they provide must be robust. > > > > > > Robust as in they may be correct here, but you lose lockdep and all > > > other infra normal mutex would give you. > > > > I know, but __netif_queue_set_napi() requires rtnl_lock() inside the > > potential > > critical section and creates a deadlock this way. However, after reading > > patches that introduce this function, I think it is called too early in the > > configuration. Seems like it should be called somewhere right after > > netif_set_real_num_rx/_tx_queues(), much later in the configuration where > > we > > already hold the rtnl_lock(). In such way, ice_vsi_rebuild() could be > > protected > > with an internal mutex. WDYT? > > On a quick look I think that may work. For setting the NAPI it makes > sense - netif_set_real_num_rx/_tx_queues() and netif_queue_set_napi() > both inform netdev about the queue config, so its logical to keep them > together. I was worried there may be an inconveniently placed > netif_queue_set_napi() call which is clearing the NAPI pointer. > But I don't see one. >
Ok, will do this in v2. Thanks for the discussion. > > > > A prettier way of protecting the same critical sections would be > > > > replacing > > > > ICE_CFG_BUSY around ice_vsi_rebuild() with rtnl_lock(), this would > > > > eliminate > > > > locking code from .ndo_bpf() altogether, ice_rebuild_pending() logic > > > > will have > > > > to stay. > > > > > > > > At some point I have decided to avoid using rtnl_lock(), if I do not > > > > have to. I > > > > think this is a goal worth pursuing? > > > > > > Is the reset for failure recovery, rather than reconfiguration? > > > If so netif_device_detach() is generally the best way of avoiding > > > getting called (I think I mentioned it to someone @intal recently). > > > > AFAIK, netif_device_detach() does not affect .ndo_bpf() calls. We were > > trying > > such approach with idpf and it does work for ethtool, but not for XDP. > > I reckon that's an unintentional omission. In theory XDP is "pure > software" but if the device is running driver will likely have to > touch HW to reconfigure. So, if you're willing, do send a ndo_bpf > patch to add a detached check. This does not seem that simple. In cases of program/pool detachment, .ndo_bpf() does not accept 'no' as an answer, so there is no easy existing way of handling !netif_device_present() either. And we have to notify the driver that pool/program is no longer needed no matter what. So what is left is somehow postpone pool/prog removal to after the netdev gets attached again.
