Re: [Qemu-devel] [PATCH 3/3] iscsi_truncate: ensure there are no requests in flight

[Kevin Wolf <kwolf@redhat.com>]

Am 11.03.2013 um 12:52 hat Peter Lieven geschrieben:
> On 11.03.2013 12:44, Paolo Bonzini wrote:
> >Il 11/03/2013 11:19, Peter Lieven ha scritto:
> >>
> >>Am 11.03.2013 um 11:16 schrieb Paolo Bonzini <pbonzini@redhat.com>:
> >>
> >>>Il 11/03/2013 11:05, Peter Lieven ha scritto:
> >>>>ensure that there are no pending I/Os before calling
> >>>>the sync readcapacity commands. the block_resize monitor
> >>>>command will also flush all I/O, but double check in
> >>>>case iscsi_truncate() is called from elsewhere in the
> >>>>future.
> >>>>
> >>>>Signed-off-by: Peter Lieven <pl@kamp.de>
> >>>>---
> >>>>block/iscsi.c |    4 ++++
> >>>>1 file changed, 4 insertions(+)
> >>>>
> >>>>diff --git a/block/iscsi.c b/block/iscsi.c
> >>>>index 3d52921..de20d53 100644
> >>>>--- a/block/iscsi.c
> >>>>+++ b/block/iscsi.c
> >>>>@@ -1167,6 +1167,10 @@ static int iscsi_truncate(BlockDriverState *bs,
> >>>>int64_t offset)
> >>>>         return -ENOTSUP;
> >>>>     }
> >>>>
> >>>>+    /* ensure all async requests are completed before executing
> >>>>+     * a sync readcapacity */
> >>>>+    bdrv_drain_all();
> >>>>+
> >>>>     if ((ret = iscsi_readcapacity_sync(iscsilun)) != 0) {
> >>>>         return ret;
> >>>>     }
> >>>
> >>>NACK to this patch.  It would be a bug, let's fix it properly.
> >>
> >>ok, are you sure that never ever will some new feature call bdrv_truncate() on an iscsi
> >>device?
> >
> >I'm sure some new feature _will_ call bdrv_truncate().  But as things
> >stand now, it would be a bug to do so on an iSCSI device.
> >
> >For example, qcow1 (and VHDX) will already call it, but that's a bug and
> >should be fixed otherwise.  Your patch will just cause an assertion failure.
> 
> In which case can qcow1 (and VHDX) be used in conjunction with an iSCSI Target?
> 
> Would it be ok for you to assert if iscsi_truncate() is called when there
> are aio's in flight? How can this be checked?
> 
> I just want to make sure that nothing nasty happens in that case (such as
> nested event loops).

Isn't the real problem that I/O requests for _this_specific_ iscsi BDS
must not be in flight? So what you reall need is bdrv_drain(iscsi_bs)?

If I understand the code correctly, this boils down to:

    while (iscsi_process_flush(iscsilun)) {
        qemu_aio_wait();
    }

Kevin