On 12/12/2016 9:07 PM, Cong Wang wrote:
On Mon, Dec 12, 2016 at 8:04 AM, Shahar Klein <shah...@mellanox.com> wrote:On 12/12/2016 3:28 PM, Daniel Borkmann wrote:Hi Shahar, On 12/12/2016 10:43 AM, Shahar Klein wrote:Hi All, sorry for the spam, the first time was sent with html part and was rejected. We observed an issue where a classifier instance next member is pointing back to itself, causing a CPU soft lockup. We found it by running traffic on many udp connections and then adding a new flower rule using tc. We added a quick workaround to verify it: In tc_classify: for (; tp; tp = rcu_dereference_bh(tp->next)) { int err; + if (tp == tp->next) + RCU_INIT_POINTER(tp->next, NULL); We also had a print here showing tp->next is pointing to tp. With this workaround we are not hitting the issue anymore. We are not sure we fully understand the mechanism here - with the rtnl and rcu locks. We'll appreciate your help solving this issue.Note that there's still the RCU fix missing for the deletion race that Cong will still send out, but you say that the only thing you do is to add a single rule, but no other operation in involved during that test?Hmm, I thought RCU_INIT_POINTER() respects readers, but seems no? If so, that could be the cause since we play with the next pointer and there is only one filter in this case, but I don't see why we could have a loop here.Do you have a script and kernel .config for reproducing this?I'm using a user space socket app(https://github.com/shahar-klein/noodle)on a vm to push udp packets from ~2000 different udp src ports ramping up at ~100 per second towards another vm on the same Hypervisor. Once the traffic starts I'm pushing ingress flower tc udp rules(even_udp_src_port->mirred, odd->drop) on the relevant representor in the Hypervisor.Do you mind to share your `tc filter show dev...` output? Also, since you mentioned you only add one flower filter, just want to make sure you never delete any filter before/when the bug happens? How reproducible is this?
The bridge between the two vms is based on ovs and representors. We have a dpif in the ovs creating tc rules from ovs rules. We set up 5000 open flow rules looks like this: cook......, udp,dl_dst=24:8a:07:38:a2:b2,tp_src=7000 actions=drop cook......, udp,dl_dst=24:8a:07:38:a2:b2,tp_src=7002 actions=drop cook......, udp,dl_dst=24:8a:07:38:a2:b2,tp_src=7004 actions=drop . . .and then fire up 2000 udp flows starting at udp src 7000 and ramping up at 100 flows per second so after 20 seconds we suppose to have 2000 active udp flows and half of them are dropped at the tc level.
The first packet of any such match hits the miss rule in the ovs datapath and pushed up to the user space ovs which consult the open flows rules above and translate the ovs rule to tc rule and push the rule back to the kernel via netlink. I'm not sure I understand what happens to the second packet of the same match or all the following packets in the same match till the tc datapath is 'ready' for them.
The soft lockup is easily reproducible using this scenario but it won't happen if we use a much more easy traffic scheme first, say 100 udp flows at 3 per second.
I added a print and a panic when hitting the loop(output attached) Also attached our .config
Thanks!
.config.gz
Description: application/gzip
tc_classify_panic.gz
Description: application/gzip
