Re: [Qemu-block] [Qemu-devel] [PATCH RFC 2/2] tests/virtio-blk: add test for WRITE_ZEROES command

[Michael S. Tsirkin]

On Fri, Jan 25, 2019 at 03:12:45PM +0000, Stefan Hajnoczi wrote:
> On Fri, Jan 25, 2019 at 09:49:03AM +0100, Thomas Huth wrote:
> > On 2019-01-25 09:16, Stefano Garzarella wrote:
> > > On Fri, Jan 25, 2019 at 07:07:35AM +0100, Thomas Huth wrote:
> > >> On 2019-01-25 07:01, Thomas Huth wrote:
> > >>> On 2019-01-24 18:23, Stefano Garzarella wrote:
> > >>>> If the WRITE_ZEROES feature is enabled, we check this
> > >>>> command in the test_basic().
> > >>>>
> > >>>> Signed-off-by: Stefano Garzarella <address@hidden>
> > >>>> ---
> > >>>>  tests/virtio-blk-test.c | 63 +++++++++++++++++++++++++++++++++++++++++
> > >>>>  1 file changed, 63 insertions(+)
> > >>>>
> > >>>> diff --git a/tests/virtio-blk-test.c b/tests/virtio-blk-test.c
> > >>>> index 04c608764b..8cabbcb85a 100644
> > >>>> --- a/tests/virtio-blk-test.c
> > >>>> +++ b/tests/virtio-blk-test.c
> > >>>> @@ -231,6 +231,69 @@ static void test_basic(QVirtioDevice *dev, 
> > >>>> QGuestAllocator *alloc,
> > >>>>  
> > >>>>      guest_free(alloc, req_addr);
> > >>>>  
> > >>>> +    if (features & (1u << VIRTIO_BLK_F_WRITE_ZEROES)) {
> > >>>> +        struct virtio_blk_discard_write_zeroes *dwz_hdr;
> > >>>> +        void *expected;
> > >>>> +
> > >>>> +        /*
> > >>>> +         * WRITE_ZEROES request on the same sector of previous test 
> > >>>> where
> > >>>> +         * we wrote "TEST".
> > >>>> +         */
> > >>>> +        req.type = VIRTIO_BLK_T_WRITE_ZEROES;
> > >>>> +        req.data = g_malloc0(512);
> > >>>
> > >>> Wouldn't it be more interesting to do a memset(req.data, 0xaa, 512) or
> > >>> something similar here, to see whether zeroes or 0xaa is written?
> > >>
> > >> Ah, never mind, I thought req.data would be a sector buffer here, but
> > >> looking at the lines below, it apparently is something different.
> > >>
> > >> Why do you allocate 512 bytes here? I'd rather expect
> > >> g_malloc0(sizeof(struct virtio_blk_discard_write_zeroes)) here. ... and
> > >> then you could also use a local "struct virtio_blk_discard_write_zeroes
> > >> dwz_hdr" variable instead of a pointer, and drop the g_malloc0() 
> > >> completely?
> > >>
> > > 
> > > Hi Thomas,
> > > it was my initial implementation, but on the first test I discovered
> > > that virtio_blk_request() has an assert on the data_size and it requires
> > > a multiple of 512 bytes.
> > > Then I looked at the virtio-spec #1, and it seems that data should be
> > > multiple of 512 bytes also if it contains the struct
> > > virtio_blk_discard_write_zeroes. (I'm not sure)
> > > 
> > > Anyway I tried to allocate only the space for that struct, commented the
> > > assert and the test works well.
> > > 
> > > How do you suggest to proceed?
> > 
> > Wow, that's a tough question. Looking at the virtio spec, I agree with
> > you, it looks like struct virtio_blk_discard_write_zeroes should be
> > padded to 512 bytes here. But when I look at the Linux sources
> > (drivers/block/virtio_blk.c), I fail to see that they are doing the
> > padding there (but maybe I'm just too blind).
> 
> The only evidence for "pad to 512 bytes" interpretation that I see in
> the spec is "u8 data[][512];".  Or have I missed something more
> explicit?

That's it. But if it doesn't mean "any number of 512 byte chunks"
then what does it mean?

> Based on the Linux guest driver code and the lack of more evidence in
> the spec, I'm pretty sure data[] doesn't need to be padded to 512 bytes
> for discard/write zero requests.

OK. Must devices support such padding?

> > Looking at the QEMU sources, it seems like it can deal with both and
> > always sets the status right behind the last byte:
> > 
> >     req->in = (void *)in_iov[in_num - 1].iov_base
> >               + in_iov[in_num - 1].iov_len
> >               - sizeof(struct virtio_blk_inhdr);
> > 
> > Anyway, I think the virtio spec should be clearer here to avoid bad
> > implementations in the future, so maybe Changpeng or Michael could
> > update the spec here a little bit?
> 
> Yep, good point.  VIRTIO 1.1 is available for public comments, so I've
> CCed the list.
> 
> Stefan

Thanks!
Care creating a github issue? And maybe proposing a spec patch.

> >  Thomas
> > 
> > 
> > > [1](https://github.com/oasis-tcs/virtio-spec/blob/master/content.tex#L3944)
> > > 
> > > 
> > > Thanks,
> > > Stefano
> > > 
> > 
> >