On Sat, 2011-02-05 at 13:45 +0000, Anton Ivanov wrote: > Ben Hutchings wrote: > > On Mon, 2011-01-31 at 11:16 +0000, Anton Ivanov wrote: > > > >> Package: linux-2.6 > >> Version: 2.6.32-30~bpo50+1 > >> Severity: normal > >> > >> > >> I keep getting VM failure messages. I suspect the machine > >> is simply a bit too slow for the network card which is in > >> it. It is a via Nehemia at 1.7GHz with an extra Intel > >> GigE server adapter. The backtraces look like showing > >> problems in the network receive/xmit routines. > >> > > > > This is an allocation failure for a *huge* allocation (order 5 = 128 KB > > chunk) in atomic (non-sleeping) context. I think this may be related to > > (1) use of GRO on the receive path to coalesce packets (2) a > > netfilter/iptables rule that requires the packet to be duplicated, or > > requires the contents to be made contiguous. > > > > 1. Do you mean gso? I do not see gro as an option on ethtool. I mean what I said. Install ethtool from squeeze. > 2. I think I know the culprit. I have recently made the machine to > double up as a X-term. Some pixmap updates can easily pass around chunks > that size. I have a couple of other systems with similar hardware so I > will see if I can reproduce it with them. That doesn't require contiguous blocks. But it will still reduce the amount of free memory. > 3. While the machine has a few netfilter rules they are all on another > interface (towards a wifi AP) and it does not do any NAT so no need to > reconstruct packets. That's strange. > >> The machine is swapless and is used mostly as an NFS > >> server. It was not showing this behaviour under 2.6.26 > >> > > [...] > > > > Probably because e1000 did not use LRO or GRO there. You can test this > > by turning off GRO with 'ethtool -K eth0 gro off'. > > > > However I would also recommend configuring the machine with some swap > > space. The kernel has trouble defragmenting memory without swapping. > > > It is my always-on server with everything raid-ed. If I configure swap > the reliability is out of the window. I did that mistake once a while > back (7 years or so) and it ended up with some serious damage. The only > to get swap for it is hardware RAID. Really, you think Linux hasn't improved in 7 years? Ben. -- Ben Hutchings Once a job is fouled up, anything done to improve it makes it worse.
Attachment:
signature.asc
Description: This is a digitally signed message part