https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=268910
Bug ID: 268910
Summary: ixgbe(4): rxcsum register setting can cause TCP
connection hangs
Product: Base System
Version: 12.3-RELEASE
Hardware: amd64
OS: Any
Status: New
Severity: Affects Only Me
Priority: ---
Component: kern
Assignee: [email protected]
Reporter: [email protected]
Hello,
I started diagnosing a FreeBSD-12.3 system with a X550-T2 10GbE controller that
was having trouble making outgoing TCP connections (ssh for example). I used
tcpdump and could see a response returning to the host but the client software
didn't seem to receive the packet and move forward.
The port was configured for 1 queue per direction
(dev.ix.0.iflib.override_nrxqs=1 and dev.ix.0.iflib.override_ntxqs=1) and all
offloads were disabled. I discovered that enabling rxcsum restored stable
networking. I could complete ssh connections as well as generic TCP/UDP
connections using nc. As soon as I removed the rxcsum offload, the initial
behavior returned. I also discovered setting the number of queues > 1 was
stable while queues=1 hung connections.
I looked at ixgbe_initialize_receive_units() in if_ix.c where the RXCSUM
register is set. Here's the relevant section:
rxcsum = IXGBE_READ_REG(hw, IXGBE_RXCSUM);
ixgbe_initialize_rss_mapping(sc);
if (sc->num_rx_queues > 1) {
/* RSS and RX IPP Checksum are mutually exclusive */
rxcsum |= IXGBE_RXCSUM_PCSD;
}
if (ifp->if_capenable & IFCAP_RXCSUM)
rxcsum |= IXGBE_RXCSUM_PCSD;
/* This is useful for calculating UDP/IP fragment checksums */
if (!(rxcsum & IXGBE_RXCSUM_PCSD))
rxcsum |= IXGBE_RXCSUM_IPPCSE;
IXGBE_WRITE_REG(hw, IXGBE_RXCSUM, rxcsum);
Not hard to see that if either queues>1 OR rxcsum is enabled the PCSD bit is
set and IPPCSE is not set. In the case where queues==1 AND !rxcsum, PCSD is not
set and IPPCSE is set (and my TCP connections hang).
I believe there is a problem here but I'm not sure which part. A few notes:
* The linux driver unconditionally sets PCSD and doesn't touch IPPCSE in
ixgbe_set_mrqc().
* The e1000 FreeBSD driver sets PCSD if rx_queues > 1. It only calls
xxx_initialize_rss_mapping() if rx_queues > 1 is true. ixgbe calls
xxx_initialize_rss_mapping outside of the conditional so it is always run. Does
this mean PCSD must always be set?
* I couldn't find anywhere in the driver where the UDP/IP fragment checksum was
used. Is there any benefit to setting IPPCSE?
I modified the above code to always set PCSD and never set IPPCSE. In my
testing on the X550-T2 and an optical X520 82599ES, this change eliminated the
hangs even with queues=1 and -rxcsum.
I'm hoping this is enough detail for someone to identity the true root cause
and best solution. Please let me know if you need additional information.
--
You are receiving this mail because:
You are the assignee for the bug.
