Skip to contentSkip to conversation

distributor: exiting workers can hang distributor core

dpdk-dev [email protected] 1 participant 1 message Sep 2, 2021

lore
Share & tools
lore
#0 lore raw

https://bugs.dpdk.org/show_bug.cgi?id=804

            Bug ID: 804
           Summary: distributor: exiting workers can hang distributor core
           Product: DPDK
           Version: 21.08
          Hardware: All
                OS: All
            Status: UNCONFIRMED
          Severity: normal
          Priority: Normal
         Component: examples
          Assignee: [email protected]
          Reporter: [email protected]

Target Milestone: ---

Created attachment 171
--> https://bugs.dpdk.org/attachment.cgi?id=171&action=edit
patch to avoid hanging

Hello,

I have been testing the distributor example using one interface and five cores
- ./build/distributor_app -l 10-14 -- -p 1. This leaves me one worker thread. I noticed the application often hangs after I supply SIGINT and I have to manually kill the process.

I added some additional debugging and discovered the distributor core is not returning from rte_distributor_flush(). Looking inside that function, I see the loop waiting for total_outstanding() to return 0. I believe that requires all workers to have returned all in-flight packets?

Moving to the worker core, it calls rte_distributor_get_pkt(), does its processing, then loops to the start of the while(!quit_signal_work) block. When exiting, this is true, so it immediately exits the loop - without returning the last batch of packets. I believe this is what is causing the distributor core to fail to exit as it continues to loop waiting for workers to return their buffers.

As a test, I added a call to rte_distributor_return_pkt(d, id, buf, num) outside the worker while() loop but before the thread exits. I've run many tests and have not seen the process hang once.

Signature · 2 lines
You are receiving this mail because:
You are the assignee for the bug.