Re: [Qemu-devel] [PATCH 17/18] qemu-io: Add background write

[Fam Zheng <famz@redhat.com>]

On Thu, 09/21 16:40, Max Reitz wrote:
> On 2017-09-19 10:03, Fam Zheng wrote:
> > On Mon, 09/18 19:53, Max Reitz wrote:
> >> On 2017-09-18 08:46, Fam Zheng wrote:
> >>> On Wed, 09/13 20:19, Max Reitz wrote:
> >>>> Add a new parameter -B to qemu-io's write command.  When used, qemu-io
> >>>> will not wait for the result of the operation and instead execute it in
> >>>> the background.
> >>>
> >>> Cannot aio_write be used for this purpose?
> >>
> >> Depends.  I have been trained to dislike *_aio_*, so that's probably the
> >> initial reason why I didn't use it.
> >>
> >> Second, I'd have to fix aio_write before it can be used.  Currently,
> >> this aborts:
> >>
> >> echo 'qemu-io drv0 "aio_write -P 0x11 0 64M"' \
> >>     | x86_64-softmmu/qemu-system-x86_64 -monitor stdio \
> >>           -blockdev node-name=drv0,driver=null-co
> >>
> >> because aio_write_done thinks it's a good idea to use qemu-io's
> >> BlockBackend -- but when qemu-io is executed through the HMP, the
> >> BlockBackend is only created for the duration of the qemu-io command
> >> (unless there already is a BB).  So what I'd have to do is add a
> >> blk_ref()/blk_unref() there, but for some reason I really don't like that.
> > 
> > What is the reason? If it crashes it should be fixed anyway, I assume?
> 
> Because the AIO CB (aio_write_done()) continues to use qemu-io's BB --
> but in case of HMP's qemu-io, that is pretty much already gone once the
> command is done.

I can see aio_{read,write}_done accesses BB for accounting, we can probably skip
this part altogether if issued from HMP (because the BB is gone). This way you
don't need the blk_ref/unref pair.

Fam

> 
> That could be fixed, as I said, by blk_ref()ing the BB before aio_write
> returns (and then blk_unref()ing it in aio_write_done()).  However, I'm
> not even sure whether aio_write_done() is always executed in the main
> thread...
> 
> Other than that, I just have a bad feeling about adding the pair, not
> sure why.  Probably because it means having to carry a temporary BB
> around until the command is done, which is weird.  Well, it's not an
> issue permission-wise, because the qemu-io BB simply doesn't take the
> proper permissions (no, I'm not going to question the fact how it's then
> possible to even write to it, considering we have assertions that check
> whether the correct permissions have been taken...), and I can't think
> of another way.
> 
> In any case, you're right, it probably needs to be fixed anyway -- even
> if the fix is just not allowing aio_write with a temporary BB (i.e. from
> HMP).
> 
> Max
>