From mboxrd@z Thu Jan 1 00:00:00 1970 From: David Miller Subject: Re: [RFC 1/2] netvsc: reference counting fix Date: Sun, 14 Aug 2016 21:31:33 -0700 (PDT) Message-ID: <20160814.213133.1692896664092434326.davem@davemloft.net> References: <1470913137-29167-1-git-send-email-vkuznets@redhat.com> <1470913137-29167-3-git-send-email-vkuznets@redhat.com> <20160813113559.617b7c55@xeon-e3> Mime-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit Cc: netdev@vger.kernel.org, haiyangz@microsoft.com, linux-kernel@vger.kernel.org, devel@linuxdriverproject.org To: stephen@networkplumber.org Return-path: In-Reply-To: <20160813113559.617b7c55@xeon-e3> List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: driverdev-devel-bounces@linuxdriverproject.org Sender: "devel" List-Id: netdev.vger.kernel.org From: Stephen Hemminger Date: Sat, 13 Aug 2016 11:35:59 -0700 > This is how I think it should be fixed, but not tested yet. > > Subjec: netvsc: use device not module reference counts > > Fix how the cross-device reference counting is handled. When VF is > associated with the synthetic interface, the VF driver module should > still be able to be unloaded. The module unload code will callback > with NETDEV_UNREGISTER event which breaks the connection safely. > (Fixes 9f4b5ba5db4 hv_netvsc: Implement support for VF drivers on Hyper-V) > > Signed-off-by: Stephen Hemminger This might not work. It is assumed that when a netdev unregister happens, it may be done so at any point in time. Therefore that NETDEV_UNREGISTER event must eliminate any and all references to a given netdev. And that works perfectly fine right now with all existing subsystems that take references to netdevs. You'll have to add something so that a NETDEV_UNREGISTER event tears this VF down and thus releases it's reference to the synthetic device.