Re: [PATCH] block: make BlockConf.*_size properties 32-bit

[Roman Kagan <rvkagan@yandex-team.ru>]

On Mon, Mar 02, 2020 at 01:55:02PM +0300, Roman Kagan wrote:
> On Thu, Feb 13, 2020 at 04:55:44PM +0300, Roman Kagan wrote:
> > On Thu, Feb 13, 2020 at 06:47:10AM -0600, Eric Blake wrote:
> > > On 2/13/20 2:01 AM, Roman Kagan wrote:
> > > > On Wed, Feb 12, 2020 at 03:44:19PM -0600, Eric Blake wrote:
> > > > > On 2/11/20 5:54 AM, Roman Kagan wrote:
> > > > > > Devices (virtio-blk, scsi, etc.) and the block layer are happy to use
> > > > > > 32-bit for logical_block_size, physical_block_size, and min_io_size.
> > > > > > However, the properties in BlockConf are defined as uint16_t limiting
> > > > > > the values to 32768.
> > > > > > 
> > > > > > This appears unnecessary tight, and we've seen bigger block sizes handy
> > > > > > at times.
> > > > > 
> > > > > What larger sizes?  I could see 64k or maybe even 1M block sizes,...
> > > > 
> > > > We played exactly with these two :)
> > > > 
> > > > > > 
> > > > > > Make them 32 bit instead and lift the limitation.
> > > > > > 
> > > > > > Signed-off-by: Roman Kagan <rvkagan@yandex-team.ru>
> > > > > > ---
> > > > > >    hw/core/qdev-properties.c    | 21 ++++++++++++---------
> > > > > >    include/hw/block/block.h     |  8 ++++----
> > > > > >    include/hw/qdev-properties.h |  2 +-
> > > > > >    3 files changed, 17 insertions(+), 14 deletions(-)
> > > > > > 
> > > > > > diff --git a/hw/core/qdev-properties.c b/hw/core/qdev-properties.c
> > > > > > index 7f93bfeb88..5f84e4a3b8 100644
> > > > > > --- a/hw/core/qdev-properties.c
> > > > > > +++ b/hw/core/qdev-properties.c
> > > > > > @@ -716,30 +716,32 @@ const PropertyInfo qdev_prop_pci_devfn = {
> > > > > >    /* --- blocksize --- */
> > > > > > +#define MIN_BLOCK_SIZE 512
> > > > > > +#define MAX_BLOCK_SIZE 2147483648
> > > > > 
> > > > > ...but 2G block sizes are going to have tremendous performance problems.
> > > > > 
> > > > > I'm not necessarily opposed to the widening to a 32-bit type, but think you
> > > > > need more justification or a smaller number for the max block size,
> > > > 
> > > > I thought any smaller value would just be arbitrary and hard to reason
> > > > about, so I went ahead with the max value that fit in the type and could
> > > > be made visibile to the guest.
> > > 
> > > You've got bigger problems than what is visible to the guest. block/qcow2.c
> > > operates on a cluster at a time; if you are stating that it now requires
> > > reading multiple clusters to operate on one, qcow2 will have to do lots of
> > > wasteful read-modify-write cycles.
> > 
> > I'm failing to see how this is supposed to happen.  The guest will issue
> > requests bigger than the cluster size; why would it cause RMW?
> > 
> > Big logical_block_size would cause RMW in the guest if it wants to
> > perform smaller writes, but that's up to the user to take this tradeoff,
> > isn't it?
> > 
> > > You really need a strong reason to
> > > support a maximum larger than 2M other than just "so the guest can
> > > experiment with it".
> > 
> > Do I get you right that your suggestion is to cap the block size
> > property at 2MB?
> > 
> > Thanks,
> > Roman.
> 
> Ping?

Ping?

> > > > 
> > > > Besides this is a property that is set explicitly, so I don't see a
> > > > problem leaving this up to the user.
> > > > 
> > > > > particularly since qcow2 refuses to use cluster sizes larger than 2M and it
> > > > > makes no sense to allow a block size larger than a cluster size.
> > > > 
> > > > This still doesn't contradict passing a bigger value to the guest, for
> > > > experimenting if nothing else.
> > > > 
> > > > Thanks,
> > > > Roman.
> > > > 
> > > 
> > > -- 
> > > Eric Blake, Principal Software Engineer
> > > Red Hat, Inc.           +1-919-301-3226
> > > Virtualization:  qemu.org | libvirt.org