2.5.65-mm1: eth0: Transmit error, Tx status register 90
Hi,
I've been benchmarking file copy operations between 2.5.65-mm1 and 2.4.20-2.51 since I have noticed that transferring files from my 2.4.20 machine to 2.5.65-mm1 gives an steady 10MBps throughput but doing the opposite (from 2.5.65-mm1 to 2.4.20) gives me a mere 3MBps throughput.
After transferring a 256MB file using NFS and SCP, I have noticed several "eth0: Transmit error, Tx status register 90.". I have attached dmesg output, ioports, iomem, lsmod and lspci.
Any ideas on what's happing here? Why network transfers do complete in less time when sending data from 2.40.20 to 2.5.65-mm1 kernel, but take more than twice when doing the opposite?
Thanks!
Felipe
Signature · 5 lines
______________________________________________ http://www.linuxmail.org/ Now with e-mail forwarding for only US$5.95/yr Powered by Outblaze
Quoted 8 lines · Felipe Alfaro Solana
Hi,
I've been benchmarking file copy operations between 2.5.65-mm1 and 2.4.20-2.51 since I have noticed that transferring files from my 2.4.20 machine to 2.5.65-mm1 gives an steady 10MBps throughput but doing the opposite (from 2.5.65-mm1 to 2.4.20) gives me a mere 3MBps throughput.
Is it slow with both scp and NFS? Or just NFS?
If just NFS then yes, I see this too. Transferring files 2.5->2.4 over NFS is several times slower than 2.4->2.4 or 2.5->2.5. Quite repeatable.
Quoted 3 lines
After transferring a 256MB file using NFS and SCP, I have noticed several "eth0: Transmit error, Tx status register 90.". I have attached dmesg output, ioports, iomem, lsmod and lspci.
That's a transmit underrun. The PCI/memory system was not able to feed data into the NIC fast enough.
Please determine when this started. 2.5.64 would be a good kernel to test because it doesn't have the PCI changes.
Quoted 4 lines · Andrew Morton
Is it slow with both scp and NFS? Or just NFS?
If just NFS then yes, I see this too. Transferring files 2.5->2.4 over NFS is several times slower than 2.4->2.4 or 2.5->2.5. Quite repeatable.
You could be hitting the same problems I saw a few weeks back.
Short: There were a *lot* of bogus UDP packets being transmitted. Trond took a look at a tcpdump log of bad traffic and found there were all sorts of silly things in there like oversized frames etc.
I've not had time to look into this since then, but also see strange effects of this bug like failing md5sums after copying over NFS.
Dave