On Thu, Jul 3, 2014 at 6:50 PM, Paolo Bonzini <pbonz...@redhat.com> wrote: > Il 03/07/2014 12:45, Ming Lei ha scritto: > >> I think it may be needed: >> >> - following requests coming inside handle_notify(): >> req0, req1, req2-flush, req3 >> - both req0 and req1 queued >> - start to handle req2-flush >> - bdrv_co_flush() calls bdrv_co_flush(bs->file), which finally >> call raw_aio_flush() to send command, but the 1st two requests >> should have been submitted to fs before the flush action > > > No, bdrv_aio_flush is only guaranteed to flush requests that _completed_ > before it was sent. > > Since we haven't yet told the guest that req0 and req1 are complete, there > is no need to submit them before req2-flush.
You are right, and it is a bit confused to think between device and host, :-) Even with O_DIRECT, qemu can be thought as no cache. Thanks,
