CC: ovs-dev
Sorry.
On 01.11.2019 14:34, Ilya Maximets wrote:
On 31.10.2019 18:12, Ben Pfaff wrote:
On Thu, Sep 05, 2019 at 11:12:26PM +0530, Anil Kumar Koli wrote:
* We can't get rid of ofproto_mutex in do_bundle_commit(), or drop it
temporarily around flow translation (i.e. the call to
ofproto_packet_out_start()), because it might need to revert all the
flow table changes and dropping the mutex would allow other threads to
race in and make conflicting changes
But it seems that the issue happens on ofproto_packet_out_finish() and not on
ofproto_packet_out_start(). So, the flow translation should be OK under the
ofproto_mutex and revert is still possible under the mutex.
The only thing we need to take out of the mutex is real action execution by
the datapath triggered by ofproto_packet_out_finish(). Callers never check the
status of this function so it should be not so hard.
So, possible solution:
* Move ofproto_packet_out_finish() out of ofproto_mutex in all the callers:
* It's easy for handle_packet_out()
* For do_bundle_commit() we could temporary store all the ofproto_packet_out
entities and finish them later.
Am I missing something? Is there reason for ofproto_packet_out_finish() to
require ofproto_mutex?
Ben, Anil, what do you think?
Best regards, Ilya Maximets.
This does seem like a reasonable solution. Do you want to take a shot
at implementing it? I promise to review a patch much more quickly than
I responded to the thread :-(
I could try. The hard part here is that I can not reproduce the issue.
_______________________________________________
dev mailing list
d...@openvswitch.org
https://mail.openvswitch.org/mailman/listinfo/ovs-dev
