On 08/21/2012 06:14 PM, Paolo Bonzini wrote:

Il 21/08/2012 11:52, Stefan Hajnoczi ha scritto:

quoted

quoted

quoted

quoted

quoted

Using /sys/dev/block or /sys/dev/char seems easier, and lets you
retrieve the parameters for block devices too.

what do you mean with "block devices"?   Using "/dev/sda" instead of
"/dev/sg0"?

Yes.

quoted

quoted

quoted

quoted

quoted

However, I'm worried of the consequences this has for migration.  You
could have the same physical disk accessed with two different HBAs, with
different limits.  So I don't know if this can really be solved at all.

I know little about qemu migration now.  The pending scsi commands will be
saved and
transfered to remote machine when starting migration?

Passthrough is already a migration blocker if both hosts do not have
access to the same LUNs.

Yes, but requiring the exact same hardware may be too much.  I'm trying
to understand the problem better before committing to a threefold
spec/qemu/kernel change.

Cong, what is the limit that the host HBA enforces (and what is the
HBA)?  What commands see a problem?  Is it fixed by using scsi-block
instead of scsi-generic (if you can use scsi-block at all, i.e. it's not
a tape or similar device)?

I don't see real problem caused by the the queue limits actually. It's a 
bug which Stefan told me.

With scsi-generic, QEMU uses a bounce buffer for non-I/O commands to a
SCSI passthrough device, so the only problem in that case should be the
maximum segment size.  This could change in the future, but max_segments
and max_sectors should not yet be a problem.

about bounce buffer, do you meat the buffer allocated in 
scsi_send_command() of hw/scsi-generic.c?

Cong.

With scsi-block, QEMU will use read/write on the block device and the
host kernel will then obey the host HBA's block limits.  QEMU will still
use a bounce buffer for non-I/O commands to a scsi-block device, but the
payload is usually small for non-I/O commands.

Paolo

quoted

When both hosts do have access to the same LUNs it's possible to
extract the block queue limits (using sysfs) and compare them.

Today you can start QEMU with different image files on both hosts.
Migration will appear to work but the disk image on the destination
host could be junk.  This is a similar case, I don't see a problem
except that there should be a safety check (maybe at the libvirt
level) to make this safe.