Bug 198723

Summary:	data corruption with big mtu
Product:	Drivers	Reporter:	Andreas Pflug (pgadmin)
Component:	Network	Assignee:	drivers_network (drivers_network)
Status:	NEW ---
Severity:	normal	CC:	josh.dewald
Priority:	P1
Hardware:	All
OS:	Linux
Kernel Version:	4.13 and up	Subsystem:
Regression:	No	Bisected commit-id:

Description Andreas Pflug 2018-02-08 10:27:08 UTC

I've been running for quite a while two Debian9 machines with 4.9 kernel, with DRBD syncing data between them. The link is directly connected Intel X540-AT2, with default settings except for MTU bumped to 9710; driver is ixgbe V4.4.0-k.

After upgrading to Kernel 4.13 (which includes ixgbe Version 5.1.0-k), I get some dozen kernel logs from DRBD (which is the same version on 4.9 and 4.13)  stating that invalid header magics were received. tcpdump shows out-of-line data.

Testing with 4.14.7, I get the same problem. Switching MTU back to 1500 fixes the problem, so does disabling scatter-gather (which includes tso; disabling tso alone doesn't help)

Comment 1 Andreas Pflug 2018-02-12 16:04:04 UTC

Apparently, disabling scatter-gather does NOT fix the problem, seeing the same again.
large-receive-offload is off (default value), disabling gso and gro doesn't help either. So apparently only cutting mtu down to default 1500 helps.