[Babel-users] [BUG] Route "deadlocks" under load due to non-atomic kernel route updates
Kirill Smelkov
kirr at nexedi.com
Thu Jun 16 20:55:28 UTC 2016
On Thu, Jun 16, 2016 at 01:47:36PM -0700, Dave Taht wrote:
> On Thu, Jun 16, 2016 at 1:40 PM, Kirill Smelkov <kirr at nexedi.com> wrote:
> > On Thu, Jun 16, 2016 at 08:38:49AM -0700, Dave Taht wrote:
> >> On Thu, Jun 16, 2016 at 4:17 AM, Kirill Smelkov <kirr at nexedi.com> wrote:
> >> > On Wed, Jun 15, 2016 at 12:56:34PM +0200, Juliusz Chroboczek wrote:
> >> >> >> If I read you correctly, this looks like a kernel bug: incorrect
> >> >> >> invalidation of the route cache.
> >> >>
> >> >> [...]
> >> >>
> >> >> > What we have here is of another kind - it is inherent race condition
> >> >> > inside kernel
> >> >>
> >> >> Perhaps I'm confused, but it still looks like a kernel bug to me.
> >> >
> >> > Yes, it is a kernel bug. But in a sense it is so old and so widespread
> >> > that it has to be cared about in userspace - as with atomic route
> >> > updates we do not hit it.
> >> >
> >> > Also: atomic route updates are needed not only for avoiding this bug.
> >> > Another reason is: if we have routedel & routeadd pair, even after
> >> > routeadd the state of cache is correct, in the time between del & add,
> >> > if a packet destined to that route gets to the node, it hits
> >> > 'unreachable' route case.
> >> >
> >> > For usual packets it is only "packet lost" and TCP probably retransmits.
> >> > But for SYN packets, e.g. when a connection is going to be established,
> >> > ICMP error is returned which results in "host unreachable" error on
> >> > originator side.
> >>
> >> Yes this variant of the bug is still there, essentially, and it bugs me.
> >>
> >> (btw the facebook page you pointed to fixes they did was fascinating -
> >> they have "interesting problems" - like dealing with 1+m routes in
> >> their route table)
> >>
> >> one day a year, for several years now, I get sufficiently irked about
> >> the atomic update problem in babel to refresh my knowledge of netlink,
> >> hack babel all to hell, and have nothing work. I left myself a bunch
> >> more breadcrumbs last night in my hacked up babel version, as to what
> >> I tried and what it did wrong... (because I'm actually also chasing
> >> another bug which I'll put up in another message)....
> >>
> >> But:
> >>
> >> Why doing the equivalent of this (and understanding how it does it)
> >>
> >> ip -6 route add fd99::33/128 via fe80::120d:7fff:fe64:c992 dev eno1
> >> ip -6 route replace fd99::33/128 via fe80::120d:7fff:fe64:c991 dev wlp2s0
> >>
> >> is so hard for me to figure out - that I don't understand. But it
> >> seems to require completely tracing through the ip route code, and
> >> writing a decoder for the netlink packets created, to figure out why
> >> what I thought would be an equivalent for babel, and taking the week
> >> or more to do it...
> >>
> >> -- look! Squirrel!
> >
> > Dave, maybe this might help you: Wireshark (not tcpdump) has decoder for
> > netlink route packets:
> >
> > https://code.wireshark.org/review/gitweb?p=wireshark.git;a=blob;f=epan/dissectors/packet-netlink-route.c;hb=v2.1.1rc0-170-gc269684
>
> Groovy. Thank you. I did not know.
You are welcome.
> In discussing this with shemminger this morning, he pointed out there
> was a semantic difference between how routes can be replaced in ipv6
> and ipv4.
>
> At *one point* last night I thought I'd successfully got ipv6 to
> atomic replace, but it had failed on ipv4 - so I will revisit the work
> soon, brain cells and time willing.
Good luck. On the first glance it is strange IPv4 and IPv6 have
different route replacement semantics, but maybe there is a reason for
it.
I would be grateful for details when/if you have it clarified.
Thanks beforehand,
Kirill
> > so you can create a virtual netlink monitor interface - something along
> > the lines of
> >
> > modprobe nlmon
> > ip link add type nlmon
> > ip link set nlmon0 up
> >
> > ( see more details in e.g. https://patchwork.ozlabs.org/patch/259444/ )
> >
> > and see the actual packets exchanged between iproute and kernel.
> >
> > Also: there is pyroute2 (https://github.com/svinota/pyroute2) which has debug
> > decoder for netlink packets, but out of the box you have to specify packet type
> > explicitly:
> >
> > https://github.com/svinota/pyroute2/blob/master/docs/debug.rst
> >
> > Maybe you already know all this, but I decided to provide info anyway to make
> > sure it is not missed, because you mentioned it is hard for you to understand
> > what is going on underneath `ip -6 ...`
> >
> > Hope this might help,
> > Kirill
> >
> >
> >> >> Perhaps it would make sense to speak to netdev about that?
> >> >
> >> > Yes, makes sense. Though as this particular case is not present on 4.2+
> >> > kernels, people on netdev will probably has less interest to look into.
> >> >
> >> > I will see what can be done.
> >> >
> >> >> > Quagga, at least, switched to atomic updates some time ago, I think.
> >> >> >
> >> >> > http://patchwork.quagga.net/patch/1234/
> >> >>
> >> >> I see. I'm busy right now, but I'll be grateful for a patch.
> >> >
> >> > I see about this. Thanks for feedback.
> >> >
> >> >
> >> > On Wed, Jun 15, 2016 at 07:35:05PM -0700, Dave Taht wrote:
> >> >> > https://lab.nexedi.com/kirr/iproute2/blob/bd480e66/t/rtcache-torture
> >> >> > (also attached to this email)
> >> >> >
> >> >> > which reproduces the problem in several minutes just on one computer and
> >> >> > retested it locally: I can reliably reproduce the issue on pristine
> >> >> > Debian 3.16.7-ckt25-2 (on both Atom and Core2 notebooks) and on pristine
> >> >> > 3.16.35 on Atom (compiled by me, since Debian kernel team has not yet
> >> >> > uploaded 3.16.35 to Jessie).
> >> >>
> >> >> I have been running this script on four different machines for hours
> >> >> now without reproducing your bug on the 4.4 or later kernels. It does
> >> >> trigger on a 3.14 kernel. (it helps to do a killall fping6 before
> >> >> exiting!)
> >> >>
> >> >> It does not seem to be happening on 4.4 or later. At one level, I'm
> >> >> relieved - one last babel bug to worry about in openwrt (now 4.4
> >> >> based), although one of the platforms I work on is still stuck at
> >> >> 3.18, as is the 3.14 c2 (for now).
> >> >>
> >> >> At another level I still really, really, really wanted atomic updates
> >> >> in babel, and was clearing the decks to make a run at the right
> >> >> netlink stuff when I'd decided to confirm your bug existed or not in
> >> >> my kernels. :(. Weirdly demotivating.
> >> >>
> >> >>
> >> >> d at dancer:~/bin$ ssh root at pi3 uname -a
> >> >> Linux pi3 4.4.12-v7+ #892 SMP Thu Jun 2 15:41:19 BST 2016 armv7l GNU/Linux
> >> >> d at dancer:~/bin$ ssh root at pi2 uname -a
> >> >> Linux pi2 4.4.12-v7+ #892 SMP Thu Jun 2 15:41:19 BST 2016 armv7l GNU/Linux
> >> >> d at dancer:~/bin$ uname -a
> >> >> Linux dancer 4.5.0-rc7-fqfi #1 SMP PREEMPT Mon Mar 7 16:04:17 PST 2016
> >> >> x86_64 x86_64 x86_64 GNU/Linux
> >> >>
> >> >> ...
> >> >>
> >> >> The odroid C2 has the bug.
> >> >>
> >> >> d at dancer:~/bin$ ssh root at c2 uname -a
> >> >> Linux c2 3.14.29-56 #1 SMP PREEMPT Wed Apr 20 12:15:54 BRT 2016
> >> >> aarch64 aarch64 aarch64 GNU/Linux
> >> >>
> >> >> BUG: Got unexpected unreachable route for 2226:3333:4444:5555::1: #
> >> >> I'd changed the number
> >> >> unreachable 2226:3333:4444:5555::1 from :: dev lo src fd99::2 metric
> >> >> 0 \ cache error -101
> >> >>
> >> >> route table for root 2226:3333:4444::/48
> >> >> ---- 8< ----
> >> >> unicast 2226:3333:4444:5555::/64 dev dum0 proto boot scope global metric 1024
> >> >> unreachable 2226:3333:4444::/48 dev lo proto boot scope global
> >> >> metric 1024 error -101
> >> >> ---- 8< ----
> >> >>
> >> >> route for 2226:3333:4444:5555::1 (once again)
> >> >> unreachable 2226:3333:4444:5555::1 from :: dev lo src fd99::2 metric
> >> >> 0 \ cache error -101 users 1 used 3
> >> >
> >> > Dave, thanks for confirming and for feedback about this.
> >> >
> >> > Yes, 4.2+ kernels should not have this _particular_ bug, because
> >> > https://git.kernel.org/linus/45e4fd26 reworks ip6_pol_route() for above
> >> > tested case to not lock the route table twice and not to create /128
> >> > cache entries on lookup when there is a gateway.
> >> >
> >> > BUT
> >> >
> >> > Route cache for IPv6 is still there in new kernels, and sometimes cache
> >> > entries are created. E.g. this happens on PMTU exception, but also for
> >> > lookups without gateway when associated flow has FLOWI_FLAG_KNOWN_NH set
> >> > (I don't yet know what it is yet, but still):
> >> >
> >> > https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/tree/net/ipv6/route.c?id=v4.7-rc3-55-gd325ea8#n1089
> >> >
> >> > etc.
> >> >
> >> > So _related_ problems should be there. They are probably just maybe less
> >> > easily reproducible and less often happening. I have not looked into
> >> > further details though...
> >> >
> >> > And also: as shown above it is better to have atomic route updates even
> >> > without cache issues to get SYN not occasionally rejected in the time of
> >> > route update.
> >> >
> >> > So Dave, please keep up your motivation for fixing this if you were
> >> > going to eventually do so.
> >> >
> >> > Thanks,
> >> > Kirill
> >> >
> >> > P.S.
> >> >
> >> >> (it helps to do a killall fping6 before exiting!)
> >> >
> >> > There is
> >> >
> >> > trap 'kill $(jobs -p)' EXIT
> >> >
> >> > it does not work?
> >> >
> >> >
> >> >> > It is always the same: the issue reproduces reliably in several minutes.
> >> >> > And it looks like e.g.
> >> >> >
> >> >> > ----- 8< ----
> >> >> > root at mini:/home/kirr/src/tools/net/iproute2/t# time ./rtcache-torture
> >> >> > PING 2222:3333:4444:5555::1(2222:3333:4444:5555::1) 56 data bytes
> >> >> > E.E.E.....E......E..E............E...E..
> >> >> > <more output from ping>
> >> >> >
> >> >> > BUG: Linux mini 3.16.35-mini64 #14 SMP PREEMPT Sun Jun 12 19:41:09 MSK 2016 x86_64 GNU/Linux
> >> >> > BUG: Got unexpected unreachable route for 2222:3333:4444:5555::1:
> >> >> > unreachable 2222:3333:4444:5555::1 from :: dev lo src 2001:67c:1254:20::1 metric 0 \ cache error -101
> >> >> >
> >> >> > route table for root 2222:3333:4444::/48
> >> >> > ---- 8< ----
> >> >> > unicast 2222:3333:4444:5555::/64 dev dum0 proto boot scope global metric 1024
> >> >> > unreachable 2222:3333:4444::/48 dev lo proto boot scope global metric 1024 error -101
> >> >> > ---- 8< ----
> >> >> >
> >> >> > route for 2222:3333:4444:5555::1 (once again)
> >> >> > unreachable 2222:3333:4444:5555::1 from :: dev lo src 2001:67c:1254:20::1 metric 0 \ cache error -101 users 1 used 4
> >> >> >
> >> >> > real 0m49.938s
> >> >> > user 0m4.488s
> >> >> > sys 0m5.872s
> >> >> > ---- 8< ----
> >> >> >
> >> >> > The issue should not show itself with kernels >= 4.2, because there the
> >> >> > lookup procedure does not take table lock twice, and /128 cache entries
> >> >> > are not routinely created (they are created only upon PMTU exception).
> >> >> >
> >> >> > I'm running Debian testing on my development machine. Currently it has
> >> >> > 4.5.5-1 (2016-05-29). I can confirm that /128 route cache entries are
> >> >> > not created there just because a route was looked up.
> >> >> >
> >> >> > Kirill
> >> >> >
> >> >> >
> >> >> > ---- 8< ---- (rtcache-torture)
> >> >> > #!/bin/sh -e
> >> >> > # torture for IPv6 RT cache, trying to hit the race between lookup,cache-add & route add
> >> >> > # http://lists.alioth.debian.org/pipermail/babel-users/2016-June/002547.html
> >> >> >
> >> >> >
> >> >> > tprefix=2222:3333:4444 # "whole-network" prefix for tests /48
> >> >> > tsubnet=$tprefix:5555 # subnetwork for which "to" route will be changed /64
> >> >> > taddr=$tsubnet::1 # test address on $tsubnet
> >> >> >
> >> >> > # setup for tests:
> >> >> >
> >> >> > # dum0 dummy device
> >> >> > ip link del dev dum0 2>/dev/null || :
> >> >> > ip link add dum0 type dummy
> >> >> > ip link set up dev dum0
> >> >> >
> >> >> > # clean route table for tprefix with only unreachable whole-network route
> >> >> > ip -6 route flush root $tprefix::/48
> >> >> > ip -6 route add unreachable $tprefix::/48
> >> >> > ip -6 route flush cache
> >> >> >
> >> >> > ip -6 route add $tsubnet::/64 dev dum0
> >> >> >
> >> >> >
> >> >> > # put a lot of requests to rt/rtcache getting route to $taddr
> >> >> > trap 'kill $(jobs -p)' EXIT
> >> >> > rtgetter() {
> >> >> > # NOTE we cannot do this with `ip route get ...` in a loop, as `ip route
> >> >> > # get` first takes RTNL lock, and thus will be completely serialized with
> >> >> > # e.g. route add and del.
> >> >> > #
> >> >> > # Ping, like other usually connect/tx activity works without RTNL held.
> >> >> > exec ping6 -n -f $taddr
> >> >> > }
> >> >> > rtgetter &
> >> >> >
> >> >> > # do route del/route in busyloop;
> >> >> > # after route add: check route get $addr is not unreachable
> >> >> > while true; do
> >> >> > ip -6 route del $tsubnet::/64 dev dum0
> >> >> > ip -6 route add $tsubnet::/64 dev dum0
> >> >> > r=`ip -6 -d -o route get $taddr`
> >> >> > if echo "$r" | grep -q unreachable ; then
> >> >> > echo
> >> >> > echo
> >> >> > echo BUG: `uname -a`
> >> >> > echo BUG: Got unexpected unreachable route for $taddr:
> >> >> > echo "$r"
> >> >> > echo
> >> >> > echo "route table for root $tprefix::/48"
> >> >> > echo "---- 8< ----"
> >> >> > ip -6 -d -o route show root $tprefix::/48
> >> >> > echo "---- 8< ----"
> >> >> > echo
> >> >> > echo "route for $taddr (once again)"
> >> >> > ip -6 -d -o -s -s -s route get $taddr
> >> >> > exit 1
> >> >> > fi
> >> >> > done
More information about the Babel-users
mailing list