On 04/11/2012 08:27 PM, Kevin Wolf wrote: > Am 11.04.2012 12:33, schrieb Kevin Wolf: >> Am 11.04.2012 12:05, schrieb Liu Yuan: >>> On 04/11/2012 03:49 PM, Kevin Wolf wrote: >>> >>>> I don't agree with this patch. If the documentation says that qemu-img >>>> always uses writeback, then the documentation must be fixed. >>>> >>>> We really don't care about flushes during an image conversion. It should >>>> just go as fast as it can. If any error happens in the middle, you'll >>>> have to throw the image away anyway, because it contains only half of >>>> what you need. So not having flushed doesn't really make any difference. >>>> >>>> As soon as a guest begins using the image with data that we really care >>>> about, it's probably running in a VM with non-unsafe mode, and then it >>>> gets flushed. >>> >>> >>> But we can set cache mode for qemu-img convert. So at least if we set >>> explicitly with other modes(for e.g, writeback) that ask to flush the >>> dirty bits, I think we should call brdv_flush(). >>> >>> If you agree, I'll update the patch with this modification. >> >> Yes, adding the flush is fine, I just wouldn't change the default cache >> mode. > > On second thought, wouldn't make it more sense to add the flush in > bdrv_close(), so that all callers get it? >
Hmm, yes, if high level bdrv_close() calls bdrv_flush() internally, we can safely call bdrv_close() from users (qemu-img) to ensure data writeback with proper cache mode. Thanks, Yuan
