Samba crashing plug on writes.

[BinaryJay]

Recently as part of a recent series of updates that have caused all sorts of mayhem on my plugs, Samba was updated as well. The behavior now is that whenever I attempt to make a write from anywhere, Samba crashes the whole plug - it drops all connections and does not respond to a physical keyboard connected issuing Ctrl+Alt+Del. The only thing I can do is power cycle it, and it's fine again until as soon as I try to write a file via Samba. Reading from Samba works just as well as ever.

This I have tried are fsck the filesystem (did find errors, including in a samba related file that I can no longer recall). Force reinstall of Samba.

Things I have not yet tried ... downgrading Samba. Will get to that when I have time. Of course I had lost my package cache when trying to get pacman back to a friendly state after messing around with things too much during the great kmod disaster of 2012 so I haven't got any prior versions of the samba package anyway.

I access external drives throughout the network using Samba so this is really stinging. Has anybody else seen this behavior or can reproduce it? Everything was working great for ages Samba wise prior to this. There was no evidence of the problem at all in my logs, it's that bad...

[WarheadsSE]

I saw this last night. Fired up all the updates on my PPV4, unpacked the modules.dep pack, and tried a test transfer (not looking to replicate, it just happened). Dropped off the network instantly.

I will hook up a serial line at my earliest convenience

[WarheadsSE]

I got home with a little time to spare and hooked up the serial on a my V4.

Initiating a write to samba locks the device, even at the serial console, without any message or warning.

[BinaryJay]

That doesn't sound promising at all. Does downgrading the package fix it for you and if so can you please link it so I can wget it?

[WarheadsSE]

Nope, nothing helps yet.

[kmihelich]

Turns out it was the network DMA that I enabled in 3.1.10. No other network transfers seemed to be affected by the addition, but reverting solves this issue so it's gone. 3.1.10-2 is in the repo now for updating.

[BinaryJay]

Thanks for making quick work of that bug. Out of curiosity, what were you hoping to accomplish by that change originally?

[kmihelich]

It was something I ran across in the kirkwood tech specs when looking up something else unrelated. I did some research on implementation of the network DMA code for the SoC in the kernel and found some mixed messages, but they were all from quite a while ago so I thought I'd give it a shot and see if things had improved. I don't run samba here, but in general the tests I ran before putting the kernel into the repo showed gains in network performance in the range of about 20% (unencrypted), most notably in the case of iSCSI. Apparently samba trips some bug in the process that other transfers hadn't caused for me.