On Thu, Mar 27, 2014 at 11:32 AM, Jesse Gross <je...@nicira.com> wrote: > On Thu, Mar 27, 2014 at 10:44 AM, Flavio Leitner <f...@redhat.com> wrote: >> On Thu, Mar 27, 2014 at 10:37:32AM -0700, Jesse Gross wrote: >>> On Thu, Mar 27, 2014 at 10:33 AM, Flavio Leitner <f...@redhat.com> wrote: >>> > On Thu, Mar 27, 2014 at 10:19:23AM -0700, Pravin Shelar wrote: >>> >> On Thu, Mar 27, 2014 at 7:05 AM, Flavio Leitner <f...@redhat.com> wrote: >>> >> > There are two problematic situations. >>> >> > >>> >> > A deadlock can happen when is_percpu is false because it can get >>> >> > interrupted while holding the spinlock. Then it executes >>> >> > ovs_flow_stats_update() in softirq context which tries to get >>> >> > the same lock. >>> >> > >>> >> > The second sitation is that when is_percpu is true, the code >>> >> > correctly disables BH but only for the local CPU, so the >>> >> > following can happen when locking the remote CPU without >>> >> > disabling BH: >>> >> > >>> >> > CPU#0 CPU#1 >>> >> > ovs_flow_stats_get() >>> >> > stats_read() >>> >> > +->spin_lock remote CPU#1 ovs_flow_stats_get() >>> >> > | <interrupted> stats_read() >>> >> > | ... +--> spin_lock remote CPU#0 >>> >> > | | <interrupted> >>> >> > | ovs_flow_stats_update() | ... >>> >> > | spin_lock local CPU#0 <--+ ovs_flow_stats_update() >>> >> > +---------------------------------- spin_lock local CPU#1 >>> >> > >>> >> > This patch disables BH for both cases fixing the deadlocks. >>> >> >>> >> This bug is already fixed in OVS. >>> > >>> > Could you point me to the commit? I am not finding anything >>> > recent. >>> >>> This is the commit: >>> >>> commit 9d73c9cac76ba557fdac4a89c1b7eafe132b85a3 >>> Author: Pravin B Shelar <pshe...@nicira.com> >>> Date: Tue Dec 17 15:43:30 2013 -0800 >>> >>> datapath: Fix deadlock during stats update. >>> >>> I thought that I had sent it in the most recent batch of changes for >>> net but it looks like I missed it. >> >> That commit is incomplete. Look at the scenario #2 which I explain >> why it is needed to disable bh for all cpus and not just local ones. > > OK, I understand the second problem now. OVS master (which I am > currently working to cross-port to net-next) uses a different strategy > that also always disables bottom halves for a different reason. Since > I forgot to send the original patch, maybe we can just apply this one > to net instead and use the new stuff directly everywhere else. >
The ovs patch fixes locking issue. local cpu check is optimization for better latency which can be significant on large SMP system. I think we should fix lockdep rather than adding latency for ovs packet processing. But all this is going to change with NUMA stats anyways, so I do not have any problem pushing this new patch. _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
