[sheepdog] [PATCH v2 0/4] add support for object cache test

[MORITA Kazutaka]

At Wed, 14 Aug 2013 17:22:53 +0800,
Liu Yuan wrote:
> 
> On Tue, Aug 13, 2013 at 01:37:36PM +0900, MORITA Kazutaka wrote:
> > From: MORITA Kazutaka <morita.kazutaka at lab.ntt.co.jp>
> > 
> > v2:
> >  - rebased onto the devel branch
> >  - added necessary flush after _vdi_write()
> >  - call _filter_qemu_io() in _qemu_io()
> > 
> > This series enables us to test object cache by running the following
> > commands.
> > 
> >  # SHEEP_EXTRA_OPTIONS="-w size=100" ./check
> >  # SHEEP_EXTRA_OPTIONS="-w size=100" ./check -writethrough
> >  # SHEEP_EXTRA_OPTIONS="-w size=100" ./check -direct
> 
> our vdi read/write causes many false alarms because 
>  - write -w doesn't flush the cache
>  - read doesn't release the cache
> 
> Maybe we should use qemu-io which release the cache and support vected io,
> to replace all the 'vdi write/read'.

qemu-io cannot be alternative because it works as if VM starts before
doing I/Os and shutdowns after that.  Running qemu-io multiple times
doesn't emulate multiple VM I/Os correctly.

I guess the only way to go is adding "collie vdi cache flush" and
"collie vdi cache delete" appropriately to our test cases.

> 
> 'SHEEP_EXTRA_OPTIONS="-w size=100" ./check' is cumbersome.
> 
> ./check -writeback is better. I think we only need to test writeback. direct
> and writethrough only need one test script to test functionality because
> writethrough is pretty much the same as writeback, and direct is same as no
> cache.

Agreed.

However, I think adding SHEEP_EXTRA_OPTIONS is still useful to test
the journaling feature.

Thanks,

Kazutaka