[Qemu-devel] Re: NBD server for QEMU images

[Salvador Fandino <sfandino@yahoo.com>]

Avi Kivity wrote:
> Martin Guy wrote:
>>> - write tons of data to nbd device, data ends up in pagecache
>>> - memory gets low, kswapd wakes up, calls nbd device to actually write
>>> the data
>>> - nbd issues a request, which ends up on the nbd server on the same
>>> machine
>>> - the nbd server allocates memory
>>> - memory allocation hangs waiting for kswapd
>>
>> In other words, it can deadlock only if you are swapping to an nbd
>> device that is served by nbd-server running on the same machine and
>> kernel. 
> 
> No.  It is possible if you issue non-O_SYNC writes.

I have run some tests and found that it's easy to cause a deadlock just
untaring a file over an nbd device being served from localhost (using
the standard nbd-server or my own, it doesn't matter).

Another interesting finding is that when the deadlock happens, qemu-nbds
is inside a read() call, waiting for new nbd requests to arrive over the
socket, and so, not trying to allocate memory or writing to disk.

BTW, I am using Debian unstable with kernel 2.6.18-1-686

Regards,

 - Salva