Bug 206624 - igb(4) fails at fragmented mbufs
Summary: igb(4) fails at fragmented mbufs
Status: New
Alias: None
Product: Base System
Classification: Unclassified
Component: kern (show other bugs)
Version: CURRENT
Hardware: Any Any
: --- Affects Some People
Assignee: freebsd-net mailing list
Depends on:
Reported: 2016-01-25 23:15 UTC by Gleb Smirnoff
Modified: 2017-11-06 17:12 UTC (History)
5 users (show)

See Also:


Note You need to log in before you can comment on or make changes to this bug.
Description Gleb Smirnoff freebsd_committer 2016-01-25 23:15:21 UTC

There are at least two evindences that igb(4) fails to transmit fragmented mbufs. Of course, fragmented mbufs degrade performance and their appearance in the stack should be avoided, but still drivers must be capable to transmit packets, instead of failing them silently.

1) First evidence came in with some restructuring to the ARP code. It bite me, and I didn't have time to investigate, quickly fixed on the ARP side.

r288991 | glebius | 2015-10-07 06:10:26 -0700 (ср, 07 окт. 2015) | 10 lines

Fix regression from r287779, that bite me. If we call m_pullup()
unconditionally, we end up with an mbuf chain of two mbufs, which
later in in_arpreply() is rewritten from ARP request to ARP reply
and is sent out. Looks like igb(4) (at least mine, and at least
at my network) fails on such mbuf chain, so ARP reply doesn't go
out wire. Thus, make the m_pullup() call conditional, as it is
everywhere. Of course, the bug in igb(?) should be investigated,
but better first fix the head. And unconditional m_pullup() was
suboptimal, anyway.


2) Second evidence involves libalias(3). Let me quote Karim:

"I've hit a very interesting problem with ipfw-nat and local TCP traffic that has enough TCP options to hit a special case in m_megapullup(). Here is the story:

I am using the following NIC:

igb0@pci0:4:0:0:        class=0x020000 card=0x00008086 chip=0x150e8086 rev=0x01 hdr=0x00

And when I do ipfw nat to locally emitted packets I see packets not being processed in the igb driver for HW checksum. Now a quick search for m_pullup in the igb driver code will show that our igb driver expects a contiguous ethernet + ip header in igb_tx_ctx_setup(). Now the friendly m_megapullup() in alias.c doesn't reserve any space before the ip header for the ethernet header after its call to m_getcl like tcp_output.c (see m->m_data += max_linkhdr in tcp_output.c).

So the call to M_PREPEND() in ether_output() is forced to prepend a new mbuf for the ethernet header, leading to a non contiguous ether + ip. This in turn leads to a failure to properly read the IP protocol in the igb driver and apply the proper HW checksum function. Particularly this call in igb_tcp_ctx_setup(): ip = (struct ip *)(mp->m_data + ehdrlen);

To reproduce the issue I simply create a NAT rule for an igb interface and initiate a TCP connection locally going out through that interface (it should go through NAT obviously) something like:

ipfw nat 1 config igb0 reset
ipfw add 10 nat 1 via igb0

 Although you need to make sure you fill enough of the SYN packet to trigger the allocation of new memory in m_megapullup. You can do this by using enough TCP options so its filling up almost all of the 256 mbuf or make RESERVE something like 300 bytes in alias.c.

The fix I propose is very simple and faster for all drivers, including the ones that do perform a check for ether + ip to be contiguous upon accessing the IP header. If the leading space is available it doesn't allocate any extra space (as it should for most cases) but if for some reason the mbuf used doesn't have 100 bytes (RESERVE in megapullup) of free space it will reserve some at the front too. If the leading space isn't necessary then it won't cause any harm.

-Subproject commit cfe39807fe9b1a23c13f73aabde302046736fa1c
+Subproject commit cfe39807fe9b1a23c13f73aabde302046736fa1c-dirty
diff --git a/freebsd/sys/netinet/libalias/alias.c b/freebsd/sys/netinet/libalias/alias.c
index 876e958..dc424a6 100644
--- a/freebsd/sys/netinet/libalias/alias.c
+++ b/freebsd/sys/netinet/libalias/alias.c
@@ -1757,7 +1757,8 @@ m_megapullup(struct mbuf *m, int len) {
         * writable and has some extra space for expansion.
         * XXX: Constant 100bytes is completely empirical. */
 #define        RESERVE 100
-   if (m->m_next == NULL && M_WRITABLE(m) && M_TRAILINGSPACE(m) >= RESERVE)
+ if (m->m_next == NULL && M_WRITABLE(m) &&
+                 M_TRAILINGSPACE(m) >= RESERVE && M_LEADINGSPACE(m) >= max_linkhdr)
                return (m);

        if (len <= MCLBYTES - RESERVE) {
@@ -1779,6 +1780,7 @@ m_megapullup(struct mbuf *m, int len) {
                goto bad;

        m_move_pkthdr(mcl, m);
+ mcl->m_data += max_linkhdr;
        m_copydata(m, 0, len, mtod(mcl, caddr_t));
        mcl->m_len = mcl->m_pkthdr.len = len;

It would be nice if some FBSD comitter could review and hopefully add this patch to FBSD."
Comment 1 Eric Joyner freebsd_committer 2017-09-13 16:28:46 UTC
This should go back to -net, since they are the ones who generally maintain the 1G iflib drivers.
Comment 2 Kevin Bowling freebsd_committer 2017-09-13 17:45:42 UTC
Gleb, I am curious if the e1000 related part of this ticket occurs in HEAD with the recent iflib changes.  We will fix it up otherwise.
Comment 3 Gleb Smirnoff freebsd_committer 2017-09-13 17:56:49 UTC
To check that, you need artificially create fragmented mbufs on output. May be reverting r288991 would be easy way to do that.