Re: Increasing amount of cached (unflushed/un-invalidated) file pages causing kernel panics

[mlelstv%serpens.de@localhost (Michael van Elst)]

fgsdfg4t534tf3%onet.pl@localhost (Marcin Kaminski) writes:

>On both Linux and FreeBSD clients when I write single file to NFS
>mountpoint it transfers file with ~10MiB(max rate on that link), then
>drops to few MiB or even below 1 MiB. It's rather no buffering here as
>link activity leds blink without pause - some performance issues on
>server side I suppose. On mixed load usage server behaves as if
>detached for a moment.

NFS over UDP has problems when you overrun the network. Some implementations
try to handle such things, but often just re-invent TCP badly.


>Both aforementioned server side customization possibilities and client
>mount options seem not to affect a problem. vfs.nfs.iothreads tested
>with 1-6 value. -n with 1-6 values.

For the stacktrace it looks like NFS requests aren't parsed correctly
so that the assertion in m_align triggers. It's not clear if the bug
is in the NFS code, could also be in network code or the ethernet
driver. Part of that code is pretty old and makes lots of assumptions
that might not hold anymore.