Re: [Qemu-devel] [PATCH] This patch adds a new block driver : iSCSI

[Paolo Bonzini <pbonzini@redhat.com>]

On 09/23/2011 11:15 AM, Mark Wu wrote:
> I tested this patch with the following command:
> x86_64-softmmu/qemu-system-x86_64 --enable-kvm rhel54_1.img -m 1024 -net
> tap,ifname=tap0,script=no -net nic,model=virtio -sdl -drive
> file=iscsi://127.0.0.1/iqn.2011-09.com.example:server.target1/
> And I found that the whole qemu process would get freezed, not reachable
> via ping and no response on desktop if there's I/O targeted to the iscsi
> drive and the iscsi target was forcefully stopped. After checking the
> backtrace with gdb, I found the I/O thread got stuck on the mutex
> qemu_global_mutex , which was hold by the vcpu thread. It should be
> released before re-entering guest. But the vcpu thread was waiting for
> the completion of iscsi aio request endlessly, and therefore couldn't
> get chance to release the mutex. So the whole qemu process became
> unresponsive. But this problem doesn't exist with the combination of
> virtio and iscsi. Only the I/O process got hung on guest in this case.
> It's more acceptable.  I am not sure how to fix this problem.

You don't. :(  It's a problem in IDE emulation and anything else that 
uses qemu_aio_flush; luckily it's only called in a few places.

It would be the same with NFS instead of iSCSI, for example.

Otherwise, you could implement some kind of timeout+reconnect logic in 
the iSCSI driver or in libiscsi.

Paolo