Il 23/08/2013 09:57, Alex Bligh ha scritto: > > > --On 23 August 2013 09:48:42 +0200 Paolo Bonzini <pbonz...@redhat.com> > wrote: > >> As a matter of fact, 2^28 madvise system calls will take much, much >> longer than 100ms. > > Probably a stupid question, but why would you need to do one call per > page? It takes a 'size_t length' parameter.
Right now migration is done a page at a time, and so is madvise AFAIU. However, even with a larger length parameter I suspect it would alone take more than 2 cycles per page. So one way to do this could be to add a flag to "migrate" that would migrate devices only, and use shared memory in both the source and the target. There is still a problem, because we must make sure the destination doesn't write to memory (e.g. read firmware) when initializing the board, because that would overwrite the memory of the running instance. But it looks more promising than page flipping. Paolo
