[ENBD] Re: ENBD for NT

[Peter T. Breuer]

"A month of sundays ago Wang Gang wrote:"
> My classmate have implemented ENBD server on Windows 2000. But the
> implement of client is more complicated, and he was ill, so we haven't do it.

Very interesting! I will happily put a pointer to his work on the enbd
pages if you leave me a reference for it.

Yes, the client will be tricky. But you can simplify it by just letting
the kernel driver fire the requests out to the net directly, instead of
letting the client daemon dive down into the kernel to pick them up and
then send them out. The client can just restrict itself to passing down
the socket when it has opened it (I don't know what it will do about
error recovery! Owww).

By the way, did you try the newer nbd codes (2.4.20, 21, 22) under the
2.4.0 kernels? I see lots of downloads every day but have no reports
back.  Testing goes on here, of course.

I have been running xfs over raid1 over nbd over lvm. To localhost
it is quite easy to provoke a lockup (of course!), but the mirror
recovery to the nbd device is perfectly solid every time, even though it
is also over lvm and to localhost.  That indicates a buffer locking
issue between xfs and the the nbd block layer.  Could that be the
same mysterious lockup mechanism that some people have
reported on in the past ...  ? I can get out of it quite easily
by erroring the device and thus dumping a few requests. But I wonder of
maybe there has always been a buffer locking race between the fs
and the block layers in linux.

Data from a different o/s would help.

Bonnie measurements on nbd on localhost are near cache speeds in 2.4.0,
if you don't sync :-). 2.4.0 removes the bonnie file before it can be
written to disk, and then vanishes the rest of the pending writes to
it! Very funny :-). I bet that will fool a few benchmarks :-).



Peter