On 06/05/2012 07:55 PM, Christoph Hellwig wrote: > On Mon, Jun 04, 2012 at 10:42:00AM +0800, Liu Yuan wrote: >> From: Liu Yuan <tailai.ly at taobao.com> >> >> For a second thought, even QEMU can issue multiple requests to the same object >> we are guaranteed that those requests don't overlap on each other. Recovery >> code and IO code might access the same object, but the race is excluded by epoch >> so we don't need per-object lock at all for now. > > How are we guaranteed that? I'm pretty sure I debugged a problem with > the kernel AIO code a while ago where that was not the case when using > windows guests. (Using the qemu raw-posix backend, not sheepdog) > You meant the requests from Guest could be overlapped? Disk is seen by guest OS as an array of blocks, I can't imagine any sane driver issue overlapped requests to disk drive controller. Thanks, Yuan |