PktSrc rewrite introduces packet drops

Description

Current master drops significantly more packets than 2.3. Looks like it's the packet source reorg: going back to 3caecadf0a3028e1dad7f446a92ffb7560c9a734 lets drop rates go back to normal.

Reported by John Donaldson.

Environment

None

Activity

Show:
Jon Siwek
October 2, 2014, 1:55 AM

I haven't reproduced the problem exactly, but I was monitoring the behavior of -Bmain-loop for various commits and noticed a significant difference that might be due to a (unintended) difference in iosource::Manager::FindSoonest's fd_set construction – quickly hacking out my FD_Set wrapper returned things to "normal". Still in process of figuring that out, but thought I'd mention it since that change wasn't specific to the PktSrc reorg.

Jon Siwek
October 2, 2014, 2:05 AM

Oh, and another difference I think is that PktSrc isn't doing a SetIdle(false) when a packet is available?

Didn't appear related to the behavior I mentioned in the last comment, but maybe it does factor in to the packet drop problem.

Robin Sommer
October 2, 2014, 3:29 PM

Good catch, indeed. The old code did that, and from a quick look it
seems that got lost now. I'll lock more closely in a bit.

Jon Siwek
October 2, 2014, 6:17 PM

He confirmed off-list to me that the patch in topic/jsiwek/pktsrc-idle was enough to put drop rates back down, but more review couldn't hurt (now I understand how easy it is to break things when making modifications to the main loop )

Robin Sommer
October 2, 2014, 7:33 PM

I'll go ahead and merge this in.

Assignee

Robin Sommer

Reporter

Robin Sommer

Labels

None

External issue ID

None

Components

Affects versions

Priority

High
Configure