Re: [Qemu-devel] Re: [PATCH v2 2/4] Clean-up a little bit the RW related bits of BDRV_O_FLAGS. BDRV_O_RDONLY gone (and so is BDRV_O_ACCESS). Default value for bdrv_flags (0/zero) is READ-ONLY. Need to explicitly request READ-WRITE.

["Michael S. Tsirkin" <mst@redhat.com>]

On Mon, Jan 18, 2010 at 11:34:59AM +0100, Markus Armbruster wrote:
> "Michael S. Tsirkin" <mst@redhat.com> writes:
> 
> > On Sun, Jan 17, 2010 at 04:48:13PM +0200, Naphtali Sprei wrote:
> >> Instead of using the field 'readonly' of the BlockDriverState struct for passing the request,
> >> pass the request in the flags parameter to the function.
> >> 
> >> Signed-off-by: Naphtali Sprei <nsprei@redhat.com>
> >
> > Many changes seem to be about passing 0 to bdrv_open. This is not what
> > the changelog says the patch does. Better split unrelated changes to a
> > separate patch.
> >
> > One of the things you seem to do is get rid of BDRV_O_RDONLY.  Why is
> > this an improvement? Symbolic name like BDRV_O_RDONLY seems better than
> > 0.
> 
> BDRV_O_RDWR is a flag, just like BDRV_SNAPSHOT.  We don't have
> BDRV_DONT_SNAPSHOT, either.

Well, this just mirros the file access macros: we have RDONLY, WRONLY
and RDRW. I assume this similarity is just historical?

> The old code can't quite devide whether BDRV_O_RDWR is a flag, or
> whether to use bits BDRV_O_ACCESS for an access mode, with possible
> values BDRV_O_RDONLY and BDRV_O_RDWR.  I asked Naphtali to clean this
> up, and recommended to go with flag rather than access mode:
> 
>     In my opinion, any benefit in readability you might hope gain by
>     having BDRV_O_RDONLY is outweighed by the tortuous bit twiddling you
>     need to keep knowledge of its encoding out of its users.
> 
> http://lists.nongnu.org/archive/html/qemu-devel/2009-12/msg02504.html
> 
> [...]
> >> @@ -985,6 +986,7 @@ static int img_snapshot(int argc, char **argv)
> >>                  return 0;
> >>              }
> >>              action = SNAPSHOT_LIST;
> >> +            bdrv_oflags &= ~BDRV_O_RDWR; /* no need for RW */
> >
> > bdrv_oflags = BDRV_O_RDONLY would be clearer, and no need
> > for comment then?
> 
> BDRV_O_RDWR is a flag, and this is the clean way to clear it.
> 
> "bdrv_oflags = BDRV_O_RDONLY" assumes that everything but the access
> mode in bdrv_oflags is clear.  Tolerable, because the correctness
> argument is fairly local, but the clean way to do it would be
> 
>     bdrv_oflags = (bdrv_oflags & ~ BDRV_O_ACCESS) | BDRV_O_RDONLY;
> 
> That's what I meant by "tortuous bit twiddling".
> 
> [...]

Thinking about it, /* no need for RW */ comment can just go.  But other
places in code just do flags = 0 maybe they should all do &=
~BDRV_O_RDWR?  I don't really have an opinion here but I do think this
patch needs a better commit log (all it says "pass the request in the
flags parameter to the function") and be split up:
patch 1 - get rid of BDRV_O_RDONLY/BDRV_O_ACCESS
patch 2 - pass the request in the flags parameter to the function
patch 3 - any other fixups

As it is, sometimes e.g. BDRV_O_RDWR is replaced with 0 sometimes as
well, and it's hard to see why.

-- 
MST