On Wed, 11 Jun 2014, dormando wrote:

On Tue, 10 Jun 2014, Eric Dumazet wrote:

quoted

On Tue, 2014-06-10 at 21:16 -0700, dormando wrote:

quoted

Ran our udpkill util against 3.10.42 with both of your patches applied...
seems like it ran a bit longer than normally would with this test (15-20
minutes), then died:

Well, could you try a recent kernel instead ?

I can see some races and fixes are probably worth it.

$ git log --oneline v3.10.42..v3.15 net/ipv4/route.c
fbdc0ad ipv4: initialise the itag variable in __mkroute_input
0d5edc6 ipv4, route: pass 0 instead of LOOPBACK_IFINDEX to fib_validate_source()
aad8872 ipv4: add a sock pointer to dst->output() path.
9114615 ipv4: return valid RTA_IIF on ip route get
3ed66e9 net: replace __this_cpu_inc in route.c with raw_cpu_inc
0b8c7f6 ipv4: remove ip_rt_dump from route.c
4a4eb21 ipv4: remove ipv4_ifdown_dst from route.c
1e8d642 Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net
a625486 ipv4: fix counter in_slow_tot
cd0f0b9 ipv4: distinguish EHOSTUNREACH from the ENETUNREACH
2045cea net: remove unnecessary return's
f87c10a ipv4: introduce ip_dst_mtu_maybe_forward and protect forwarding path against pmtu spoofing
dcdfdf5 ipv4: fix race in concurrent ip_route_input_slow()
482fc60 ipv4: introduce new IP_MTU_DISCOVER mode IP_PMTUDISC_INTERFACE
0baf2b3 ipv4: shrink rt_cache_stat
0a7e226 ipv4: fix ineffective source address selection
734d272 ipv4: raise IP_MAX_MTU to theoretical limit
ca4c3fc net: split rt_genid for ipv4 and ipv6
2ffae99 ipv4: use next hop exceptions also for input routes
fe2c633 net: Convert uses of typedef ctl_table to struct ctl_table
6bc19fb Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net
5aad1de ipv4: use separate genid for next hop exceptions
f016229 ipv4: rate limit updating of next hop exceptions with same pmtu
387aa65 ipv4: properly refresh rtable entries on pmtu/redirect events

Newest I can realistically roll would be 3.14.6, so I just tried
that... Without your two patches, it still dies from the UDP bug.

--> Meant to say here that both *with* and *without* your two new patches
it still crashes.

Unfortunately 3.14 has a few regressions.. one is some bad CPU usage i'll
have to track down, and two something about pstore is broken, so I can't
get the trace from the crash. It's compressing now and has more of the
kernel log, but it's missing the actual panic part.

$ git log --oneline v3.14..v3.15 net/ipv4/route.c
fbdc0ad ipv4: initialise the itag variable in __mkroute_input
0d5edc6 ipv4, route: pass 0 instead of LOOPBACK_IFINDEX to fib_validate_source()
aad8872 ipv4: add a sock pointer to dst->output() path.
9114615 ipv4: return valid RTA_IIF on ip route get
3ed66e9 net: replace __this_cpu_inc in route.c with raw_cpu_inc
0b8c7f6 ipv4: remove ip_rt_dump from route.c
4a4eb21 ipv4: remove ipv4_ifdown_dst from route.c
1e8d642 Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net
2045cea net: remove unnecessary return's

No more obvious race fixes. I can try 3.15 fully vanilla but I'm having
doubts?

We have a few patches on top of this, but none of them are active at the
time of my test. I've tried removing them in the past and it did nothing
as well.

Sorry :(