Re: [libvirt] [PATCH] migrate VMs between different-endian hosts

On Mon, Apr 11, 2011 at 11:46:55AM -0600, Eric Blake wrote: > On 04/11/2011 07:36 AM, Daniel Veillard wrote: >>>>> @@ -3097,6 +3107,11 @@ qemuDomainSaveImageOpen(struct qemud_dri >>>>> } >>>>> >>>>> if (header.version> QEMUD_SAVE_VERSION) { >>>>> + /* convert endianess and try again */ >>>>> + bswap_header(&header); >>>>> + } >>>> Hum, isn't there a more reliable way to detect the change of >>>> endianness ? That's a bit fishy IMHO :-) >>> The problem is that the header should not have been written in a >>> hosts' native format. So what can go wrong? QEMUD_SAVE_VERSION is >>> '2'. Either we find 1 or 2 here and go ahead and accept it 'as-is'. >>> Otherwise anything bigger than 3 is not accepted and swapped. 3 then >>> becomes 0x03 00 00 00 and is discarded. 0x 02 00 00 00 would be >>> swapped to '2' and accepted. >> yeah, I understand, okay, > Should we be writing the header in a particular byte order, regardless > of host endianness? Or does that require bumping the header version to > 3 anyways? The thing I'm worried about is that by bumping older versions won't be able to restore the new dumps, and that could be a serious issue on shared storage. We are not introducing a new feature, so I thin we need to preserve forward compatibility. The patch as is is minimal but achieves it.