Date: Fri, 20 Mar 2015 16:03:24 -0400 From: Karim Fodil-Lemelin <fodillemlinkarim@gmail.com> To: freebsd-net@freebsd.org Subject: Re: Another fragment question / patch Message-ID: <550C7D0C.3090603@gmail.com> In-Reply-To: <550C5F6C.3080302@selasky.org> References: <550C3A62.3080403@gmail.com> <550C5F6C.3080302@selasky.org>
next in thread | previous in thread | raw e-mail | index | archive | help
On 2015-03-20 1:57 PM, Hans Petter Selasky wrote: > On 03/20/15 16:18, Karim Fodil-Lemelin wrote: >> Hi, >> >> While reading through a previous comment on this list about fragments >> I've noticed that mbuf tags aren't being copied from the leading >> fragment (header) to the subsequent fragment packets. In other words, >> one would expect that all fragments of a packet are carrying the same >> tags that were set on the original packet. I have built a simple test >> were I use ipfw with ALTQ and sent large packet (bigger then MTU) off >> that BSD machine. I have observed that the leading fragment (m0) packet >> is going through the right class although the next fragments are hitting >> the default class for unclassified packets. >> >> Here is a patch that makes things works as expected (all fragments carry >> the ALTQ tag): >> >> diff --git a/freebsd/sys/netinet/ip_output.c >> b/freebsd/sys/netinet/ip_output.c >> index d650949..7d8f041 100644 >> --- a/freebsd/sys/netinet/ip_output.c >> +++ b/freebsd/sys/netinet/ip_output.c >> @@ -1184,7 +1184,10 @@ smart_frag_failure: >> ipstat.ips_odropped++; >> goto done; >> } >> - m->m_flags |= (m0->m_flags & M_MCAST) | M_FRAG; >> + >> + m->m_flags |= (m0->m_flags & M_COPYFLAGS) | M_FRAG; >> + m_tag_copy_chain(m, m0, M_NOWAIT); >> + >> /* >> * In the first mbuf, leave room for the link >> header, then >> * copy the original IP header including options. The >> payload >> diff --git a/freebsd/sys/sys/mbuf.h b/freebsd/sys/sys/mbuf.h >> index 2efff38..6ad8439 100644 >> --- a/freebsd/sys/sys/mbuf.h >> > > Hi, > > I see your point about copying the tags. I'm not sure however that > M_COPYFLAGS is correct, because it also copies M_RDONLY, which is not > relevant for this case. Can you explain what flags need copying in > addition to M_MCAST ? Maybe we need to define these flags separately. > > Thank you! > > --HPS Hi, Arguably the M_RDONLY is added when m_copy() is called a bit later in that function. m_copym() does a shallow copy (through a call to mb_dupcl) and will set the RDONLY flag when doing so. So the fact it was copied over from M_COPYFLAGS shouldn't be a problem in terms of functionality. A similar logic applies to the M_VLANTAG since it should never be set in ip_output() (severe layer violation). I guess M_COPYFLAGS is kinda safe but not necessarily correct. In terms of appropriate behavior (whats the real purpose of M_COPYFLAGS?) my initial patch is debatable and to answer your question I would consider to copy the remaining flags: M_PKTHDR => already in there through the m_gethdr() call M_BCAST => no need to copy M_MCAST => already in there in ip_fragment() M_PROTOFLAGS M_SKIP_FIREWALL => for layer 2 fire-walling? So something like? - m->m_flags |= (m0->m_flags & M_MCAST); + m->m_flags |= (m0->m_flags & (M_MCAST | M_PROTOFLAGS)); + m_tag_copy_chain(m, m0, M_NOWAIT); Cheers! Karim.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?550C7D0C.3090603>