On Fri, 9 Mar 2018, David Miller wrote: > From: Ilpo Järvinen <ilpo.jarvi...@helsinki.fi> > Date: Fri, 9 Mar 2018 16:11:47 +0200 (EET) > > > Unfortunately I don't have now permission to publish the time-seq > > graph about it but I've tried to improve the changelog messages so > > that you can better understand under which conditions the problem > > occurs. > > It is indeed extremely unfortunate that you wish to justify a change > for which you cannot provide the supporting data at all.
Here is the time-seqno graph about the issue: https://www.cs.helsinki.fi/u/ijjarvin/linux/nonsackbugs/recovery_undo_bug.pdf First the correct CC action (wnd reduction) occurs; then bogus undo causes bursting back to the window with which the congestion losses occurred earlier; because of the burst, some packets get lost due to congestion again. The sender is actually somewhat lucky here: If only one packet would get lost instead of three, the same process would repeat for the next recovery (as cumulative ACK to high_seq condition would reoccur). -- i.