From mboxrd@z Thu Jan 1 00:00:00 1970 From: Ben Hutchings Subject: Re: [PATCH] bnx2: Use dev_kfree_skb_any() in bnx2_tx_int() Date: Fri, 1 Nov 2013 23:34:50 +0000 Message-ID: <1383348890.1737.40.camel@bwh-desktop.uk.level5networks.com> References: <20131031.002602.2030884257299947291.davem@davemloft.net> <20131101.180147.629210014672277622.davem@davemloft.net> Mime-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: QUOTED-PRINTABLE Cc: , , , , To: David Miller Return-path: In-Reply-To: <20131101.180147.629210014672277622.davem@davemloft.net> Sender: linux-kernel-owner@vger.kernel.org List-Id: netdev.vger.kernel.org On Fri, 2013-11-01 at 18:01 -0400, David Miller wrote: > From: Cong Wang > Date: Thu, 31 Oct 2013 21:19:16 -0700 >=20 > > 2013=E5=B9=B410=E6=9C=8830=E6=97=A5 =E4=B8=8B=E5=8D=889:26=E4=BA=8E= "David Miller" =E5=86=99=E9=81=93=EF=BC=9A > >> > >> We have to provide a softint compatible environment for this callb= ack > >> to run in else everything is completely broken. > >> > >> All these drivers can safely assume softirq safe locking is > >> sufficient, you're suggesting we need to take this hardirq safety = and > >> I'm really not willing to allow things to go that far. A lot of > >> effort has been expended precisely to avoid that kind of overhead = and > >> cost. > >=20 > > Alright, I am thinking to move netpoll_poll_dev() to a delayed work= =2E > =20 > What if the printk is outputting a message that will help us discover > that work queues are deadlocked? >=20 > You can't delay the message, because every layer of indirection you > add increases the possibility that the message it never seen. You > have to do it synchronously. As you've said, the ndo_start_xmit and NAPI poll operations are intende= d to be called in softirq context, so everything that interlocks with the= m will use spin_lock_bh(). Calling them from hardirq context obviously opens the possibility of a deadlock. How do you expect anyone to solve that? I think that most of the time netpoll doesn't actually call the NAPI poll function, and the driver ndo_start_xmit function doesn't take any locks, so we don't actually hit the deadlock in practice (on mainline kernels - RT is a different story). Obviously, the less machinery netpoll relies on continuing to work, the better, so it should preferably defer to sofirq context rather than workqueue context. I think this means hooking queue_process() into net_tx_action(), and then cutting out much of the rest of netpoll. Ben. --=20 Ben Hutchings, Staff Engineer, Solarflare Not speaking for my employer; that's the marketing department's job. They asked us to note that Solarflare product names are trademarked.