Re: [Qemu-devel] virtio-scsi limits

["Richard W.M. Jones" <rjones@redhat.com>]

On Tue, Apr 18, 2017 at 05:30:52PM +0200, Paolo Bonzini wrote:
> > When I changed libguestfs to use LUNs instead of targets (target=0
> > unit=i), I got a peculiar bug.  It looks like there is some kind of
> > race when enumerating the device, where /sys is populated before the
> > device is actually available.
> 
> That's not _too_ surprising because devtmpfs processes creation/deletion
> requests asynchronously.

I had a bit of a look into this, and I don't think this can be right.
At this point we're running inside the initramfs environment, where
/dev is a rootfs (a special case of either ramfs or tmpfs).

Would it still be the case that a single process that does mknod
followed by mount could not see the effect of the mknod?

The mount error (EINVAL) should indicate an invalid superblock rather
than a missing inode (if you believe the documentation, so that may be
wrong).

The sequence of events (happening in a single init process) is as
described below:

> > https://github.com/libguestfs/supermin/blob/master/init/init.c#L176
> > 
> > * fp = fopen ("/sys/block/sdab/dev", "r") succeeds at line 181
> > 
> > * we read major:minor from fp at line 230
> > 
> > * we mknod ("/dev/root",...) at line 245
> > 
> > * we mount ("/dev/root", "/root", "ext2"...) at line 262, which fails with
> >   EINVAL
> > 
> > * shortly after that, kernel messages indicate that sdab has been
> >   attached.

Thanks,

Rich.

-- 
Richard Jones, Virtualization Group, Red Hat http://people.redhat.com/~rjones
Read my programming and virtualization blog: http://rwmj.wordpress.com
Fedora Windows cross-compiler. Compile Windows programs, test, and
build Windows installers. Over 100 libraries supported.
http://fedoraproject.org/wiki/MinGW