Re: [Qemu-devel] [PATCH] block-backend: allow flush on devices with open tray

[Kevin Wolf]

Am 14.06.2016 um 18:13 hat Max Reitz geschrieben:
> On 14.06.2016 17:54, John Snow wrote:
> > 
> > 
> > On 06/14/2016 09:19 AM, Max Reitz wrote:
> >> On 10.06.2016 23:59, John Snow wrote:
> >>> If a device still has an attached BDS because the medium has not yet
> >>> been removed, we will be unable to migrate to a new host because
> >>> blk_flush will return an error for that backend.
> >>>
> >>> Replace the call to blk_is_available to blk_is_inserted to weaken
> >>> the check and allow flushes from the backend to work, while still
> >>> disallowing flushes from the frontend/device model to work.
> >>>
> >>> This fixes a regression present in 2.6.0 caused by the following commit:
> >>> fe1a9cbc339bb54d20f1ca4c1e8788d16944d5cf
> >>> block: Move some bdrv_*_all() functions to BB
> >>>
> >>> Signed-off-by: John Snow <address@hidden>
> >>> ---
> >>>  block/block-backend.c | 2 +-
> >>>  1 file changed, 1 insertion(+), 1 deletion(-)
> >>
> >> I'm still not sure we shouldn't do the same for blk_{co,aio}_flush(). I
> >> guess you exclude them here because you specifically want to fix the
> >> issue mentioned in the commit message, but then we could just make
> >> blk_flush_all() ignore an -ENOMEDIUM.
> > 
> > Yeah, I didn't investigate the full path. Just making the minimal fixes.
> > Is there a concern that this may still leave certain pathways broken
> > when the CDROM tray is open?
> > 
> > I don't know of any immediately without digging again.
> > 
> >>
> >> I personally think we should make all blk_*flush() functions use
> >> blk_is_inserted() instead of blk_is_available(). As we have discussed on
> >> IRC, there are probably not that many cases a guest can flush a medium
> >> in an open tray anyway (because the main use case are read-only
> >> CD-ROMs), and even if so, that wouldn't change any data, so even if the
> >> guest can actually flush something on an open tray, I don't think anyone
> >> would complain.
> >>
> >> Max
> >>
> > 
> > I have difficulty making pragmatic arguments when purity is at stake,
> > but I've already wandered outside of my device model, so I will defer to
> > your judgment.
> > 
> >>> diff --git a/block/block-backend.c b/block/block-backend.c
> >>> index 34500e6..d1e875e 100644
> >>> --- a/block/block-backend.c
> >>> +++ b/block/block-backend.c
> >>> @@ -1122,7 +1122,7 @@ int blk_co_flush(BlockBackend *blk)
> >>>  
> >>>  int blk_flush(BlockBackend *blk)
> >>>  {
> >>> -    if (!blk_is_available(blk)) {
> >>> +    if (!blk_is_inserted(blk)) {
> >>>          return -ENOMEDIUM;
> >>>      }
> >>>  
> >>>
> >>
> >>
> > 
> > Is this a NACK unless I attempt to address the wider design issue?
> 
> I just don't see a point in using blk_is_inserted() here but
> blk_is_available() in the other blk_*flush() functions. If
> blk_is_inserted() is correct for blk_flush(), it should be correct for
> blk_co_flush() and blk_aio_flush(), too.

I agree, if we can, we should keep the behaviour consistent between all
interfaces types (sync/AIO/coroutine, byte-based/sector-based) for the
same operation.

Eric also rightfully said that we need a test cases, so a v2 would be
good anyway.

Kevin

pgpuCsTNI_73X.pgp
Description: PGP signature