Control: fixed -1 3.4.1-1~experimental.1 On Tue, 2013-07-02 at 10:42 +0200, Udo Lembke wrote: > Package: src:linux > Version: 3.2.46-1 > Severity: important > Tags: upstream > > Dear Maintainer, > we running three nodes as ceph-cluster (0.61.3-1~bpo70+1) and sometimes (app. > once or twice a day) we got an kernel trace which seems to be related > to the Intel 10GB-NIC (ixgbe module). > The same happens before update (from: SMP Debian 3.2.41-2+deb7u2 x86_64 > GNU/Linux) and also occur with an actual self-compiled ixgbe-module (3.15.1). > The page allocation failure are not only swapper related. Also ceph-osd and > kswapd0 trigger this problem. > > To avoid the kernel-trace I use the sysctl-value "vm.min_free_kbytes = > 337920" but without luck.
You could try increasing it further...
> We had also the issue (perhaps not related to this bug) that the whole host
> freezed and the NIC flood the 10-GB-Port so that the complete switch (and
> the compled network) are freezed.
Please report only one issue in each bug report.
> Since that the ceph-cluster are isolated for test in an extra network.
> We will switch to productive mode if the kerneltraces are gone.
[...]
I believe the driver stopped making order-2 allocations in Linux 3.4,
with this change:
commit f800326dca7bc158f4c886aa92f222de37993c80
Author: Alexander Duyck <alexander.h.du...@intel.com>
Date: Sat Mar 3 02:35:52 2012 +0000
ixgbe: Replace standard receive path with a page based receive
This should fix the allocation failures you're seeing.
This might be backported to 3.2 if we have some other reason to update
the ixgbe driver. Otherwise it will not be.
Ben.
--
Ben Hutchings
Life would be so much easier if we could look at the source code.
signature.asc
Description: This is a digitally signed message part
