Promptness, queueing, diskless etc. etc.

[Jonathan Shapiro]

   Multitrack operations on IBM disks do not require an interrupt to 
   switch tracks.  My reading of the SCSI standard indicates they don't
   need extra interrupts either.  (Both SCSI and IBM-FBA disks treat
   disk reads/writes as a stream of fixed size blocks.)

   What do IDA disks require?

I think you mean IDE?

Not sure, but multitrack reads on ESDI, ST506, and XT drive families
do require multiple operations.  IDE is similar to SCSI, but...

On the PC it turns out that people implement exactly two disk drivers.

The first is commonly called 'hd', and covers IDE, XT, ST506, and ESDI
drives.  All of these present a compatible low-level hardware I/O
interface to the machine.  Because some of these drives (ST506, ESDI)
will not perform multicylinder operations the driver typically assumes
that none will.  In practice, most hardware (not ST506, but very few
people still run that) will do the right thing about switching from
one head to the next in the middle of an op on the same cylinder, but
switching cylinders won't be handled automatically.  It depends on how
compatible you want to be with old cards, and how smart a driver
author you are.

The other commonly implemented driver is the SCSI driver, because SCSI
does not present a compatible low-level I/O interface.  This driver
generally will not do multiple operations for a multicylinder or
multitrack read.

   > The problem lies in the definition of "depend."  If a domain holds a
   > start key to a dead domain, is that domain considered dead?  How does
   > one indicate "precious" domains that should *not* be considered dead
   > under these conditions?

   A start key to a dead domain is a solved problem (it acts like a zero
   data key).  Caveat Programmer.  I assume "precious" domains would be
   so marked.  I havn't thought much about what authority would be required
   to set the mark.

The difficulty is that the dead domain might be, say, the prime space
bank, if it happened that due to process migration the prime space
bank was on the dead machine.  Equally, it might be a meter near the
root of a critical tree.  At some point you have to get into
discussions of redundancy.


Jonathan