On Tue, Sep 19, 2000 at 07:49:51PM -0400, Drew Northup wrote:
> Great thought..., but once again Intel has broken it before you could even
> try. The main problem with the multiprocessor argument is not just the
> ring0 code (which is a real pain) it is also that SMP was designed with the
> intention that you never know what process is running on which processor,
> and thus the user level stuff doesn't know any "better" as to what it is
> really doing at the hardware level. OH well, maybe later.......;^)
>
> As for using a hard drive instead of a file, you have one MAJOR and
> undeniable sticking point.....partition tables & hard drive access
> permissions. In LINUX and other UNIX-like systems, you must mount a drive
> manually. However, in many, many other systems, a hard drive partitions of
> a matching type are AUTOMATICALLY mounted on boot. Thus you'd have two
> systems writing to a filesystem that they can both see AT THE "HARDWARE"
> LEVEL and there is no host, no master, no "in charge" system controlling
> disk access. This leads to unbelievable data loss and other catastrophic
> problems. If you'd like proof of this, set up a joint scsi bus between two
> machines, format a partition on a drive as something that they can both
> access as something that they can both access, and then make sure that they
> can both mount the partition. Then try to write to the drive from both
> machines simultaneously. The drive will become unrecoverably corrupted.
> Granted, we wouldn't be trying to do exactly the same thing, but I think
> that I made my point that the logistics of the issue of RAW disk access by
> the guest are just way beyond what we would even want to consider right now.
So you do like dosemu and "strongly suggest" unmounting the drive first. Or,
if you're really keen at hacking code, write a filesystem access layer. Sounds
quite a bit too hard to work on right now, though. As for the other systems,
there is usually some kind of locking mechanism to at least prevent fsck-type
utilities from mucking around with it at the same time (which would be the only
real cause of concern if you find a way to tell the OS to keep its buffers
flushed) -- I'm thinking Windows here; Norton Disk Doctor somehow puts a "lock"
on the drive and Speed Disk complains about that. Considering how much other
stuff you can muck around with in Windows, I wouldn't be at all suprised to
find a way to do all fs ops async, disable all disk caching, and/or totally
block access to a partition for a certain length of time. Of course there
will always be the user who thinks that they can run Linux under Linux using
raw access to the same partition... oops.
Kenneth
