Re: QEMU RBD is slow with QCOW2 images

[Kevin Wolf]

Am 04.03.2021 um 15:08 hat Stefano Garzarella geschrieben:
> On Thu, Mar 04, 2021 at 01:05:02PM +0100, Kevin Wolf wrote:
> > Am 03.03.2021 um 18:40 hat Stefano Garzarella geschrieben:
> > > Hi Jason,
> > > as reported in this BZ [1], when qemu-img creates a QCOW2 image on RBD
> > > writing data is very slow compared to a raw file.
> > > 
> > > Comparing raw vs QCOW2 image creation with RBD I found that we use a
> > > different object size, for the raw file I see '4 MiB objects', for
> > > QCOW2 I
> > > see '64 KiB objects' as reported on comment 14 [2].
> > > This should be the main issue of slowness, indeed forcing in the code 4 
> > > MiB
> > > object size also for QCOW2 increased the speed a lot.
> > > 
> > > Looking better I discovered that for raw files, we call rbd_create() with
> > > obj_order = 0 (if 'cluster_size' options is not defined), so the default
> > > object size is used.
> > > Instead for QCOW2, we use obj_order = 16, since the default 'cluster_size'
> > > defined for QCOW2, is 64 KiB.
> > 
> > Hm, the QemuOpts-based image creation is messy, but why does the rbd
> > driver even see the cluster_size option?
> > 
> > The first thing qcow2_co_create_opts() does is splitting the passed
> > QemuOpts into options it will process on the qcow2 layer and options
> > that are passed to the protocol layer. So if you pass a cluster_size
> > option, qcow2 should take it for itself and not pass it to rbd.
> > 
> > If it is passed to rbd, I think that's a bug in the qcow2 driver.
> 
> IIUC qcow2 properyl remove it, but when rbd uses qemu_opt_get_size_del(opts,
> BLOCK_OPT_CLUSTER_SIZE, 0) the default value of qcow2 format is returned.
> 
> Going in depth in qemu_opt_get_size_helper(), I found that qemu_opt_find()
> properly returns a NULL pointer, but then we call find_default_by_name()
> that returns the default value of qcow2 format (64k).

Ugh, I see why. We're passing the protocol driver a QemuOpts that was
created for a QemuOptsList with the qcow2 default, not for its own
QemuOptsList. This is wrong.

Note that the QemuOptsList is not qcow2_create_opts itself, but a list
that is created with qemu_opts_append() to combine qcow2 and rbd options
into a new QemuOptsList. For overlapping options, the format wins.

I don't think you can change the QemuOptsList of an existing QemuOpts,
nor is there a clone operation that could just copy all options into a
new QemuOpts created for the rbd QemuOptsList, so maybe the easiest
hack^Wsolution would be converting to QDict and back...

> > > Using '-o cluster_size=2M' with qemu-img changed only the qcow2 cluster
> > > size, since in qcow2_co_create_opts() we remove the 'cluster_size' from
> > > QemuOpts calling qemu_opts_to_qdict_filtered().
> > > For some reason that I have yet to understand, after this deletion, 
> > > however
> > > remains in QemuOpts the default value of 'cluster_size' for qcow2 (64 
> > > KiB),
> > > that it's used in qemu_rbd_co_create_opts()
> > 
> > So it seems you came to a similar conclusion. We need to find out where
> > the 64k come from and just fix that so that rbd uses its default.
> 
> Yes, I tried debugging above, but I'm not sure how to fix it.
> 
> Maybe a new parameter in qemu_opt_get_size_helper() to prevent it from
> looking for the default value.
> Or we should prevent the default value from being added to the
> opts->list->desc, but that part is still not very clear to me.

opts->list is already wrong, I think this is what we need to fix.

> > > At this point my doubts are:
> > > Does it make sense to use the same cluster_size as qcow2 as object_size in
> > > RBD?
> > > If we want to keep the 2 options separated, how can it be done? Should we
> > > rename the option in block/rbd.c?
> > 
> > My lazy answer is that you could just use QMP blockdev-create, where you
> > create layer by layer separately.
> > 
> > What could possibly be done for the QemuOpts is using the dotted syntax
> > like for opening, so you could specify file.cluster_size=... for the
> > protocol layer (or data_file.cluster_size=... for the external data
> > file etc.)
> 
> This would be cool :-)

I'm almost sure that compatibility will make this more complicated than
it sounds, but we could have a try.

Kevin