On 10/09/2012 04:31 PM, Stefan Hajnoczi wrote: > On Tue, Oct 09, 2012 at 10:27:26AM +0800, Dave Young wrote: >> On 10/05/2012 04:14 PM, Stefan Hajnoczi wrote: >> >>> On Sun, Sep 23, 2012 at 10:37:09AM +0800, Dave Young wrote: >>>> For the serial number decreasing issue, I think there's only these two >>>> ways to >>>> select, there's no ideal way to resolve this issue. >>>> My use case for this is for the kdump kernel to find proper disks, >>>> after 1st kernel crashing 2nd kernel need find right disk to dump vmcore. >>>> In this case v1 and v2 aproaches are both find to me. >>>> >>>> From my point of view, patch v1 is better though, I think unpluging 100000 >>>> is >>>> not a sane use case. It's not likely to happen. >>> >>> I'm not sure auto-assigning serial numbers is a good idea. The guest can >>> use >>> the serial number in /etc/fstab or other places where it expects the serial >>> number to be persistent. >>> >>> Your patch does not provide persistent serial numbers, so a change to the >>> QEMU >>> invocation could result in different serial numbers. The guest will get >>> confused or perhaps refuse to boot. >> >> >> Yes, it introduce confusion, but in this way at least the serial number >> can be persistent across guest reboot. Traditionally ide disks use this >> way as well, such as QEMU_HARDISK_00001, I think guest should not use >> this in /etc/fstab. > > If you don't want to set a persistent serial number, use another mechanism to > identify the disk. For example, Linux has /dev/disk/by-path/ which identifies > virtio-blk PCI adapters, IDE, SCSI disks, etc.
> > Does this work for your use case? I have tried this before, but after rebooting (kexec/kdump) the by-path link was not created. It might be udev bug anyway, I'm not sure though. > > Stefan > -- Thanks Dave