🚀 go-pugleaf

   In article <1992Dec18.175402.27841@hubcap.clemson.edu>
   dfk@wildcat.dartmouth.edu (David Kotz) writes:

   I have a question about DMA devices, specifically for networks and disks.

   Do you know if it is possible (and if so, common) for a DMA device to
   support scatter/gather? In particular, I'm trying to model a multiprocessor
   interconnect network interface, and it would be very helpful if I could
   avoid packetization by specifying the addresses and lengths of my message
   header and message body, for the network DMA to pick up. I'm guessing that
   this is not usually an option.

   Also, in a disk interface, can it read in a whole track but scatter the
   blocks to different locations in memory? I'm guessing that this kind of
   thing is more common.

Well, as a historical perpective on this subject, consider what the I/O
adapters of the Vax11/780 (ca. 1978?) did/do: They (Unibus and Massbus
adapters) had a set of socalled mapping registers. These mapped an offset in
an I/O transaction into the appropriate _physical_ address. A set of system
routines was provided for drivers to load the mapping registers from the page
table of a process (or system space) with the required data.  (I think there
are seperate routines to make sure the page tables contain valid entries in
the relevant page range, which is especially important for I/O directly
to/from user space. I do know there was an optimization for reads which didn't
bother to page in pages from backing store which were going to be overwritten
shortly by a read  it just substituted a page from the free list and marked
the page table entry as dirty.)

A number of smaller Vaxen (only the uVaxI? or also the 725/730?) didn't have
these and in general had to do I/O a page at a time...probably still slightly
faster than a PC/XT :).

A later adapter, the one for the CI (computer interconnect, four cables
running at 70Mbit/s each) went even further: It reads I/O request packets from
system virtual memory, containing pointers to buffers in virtual memory (I
think user space buffers are double mapped into system space), and reads the
page tables itself in order to determine physical addresses.  Sort of like a
closely coupled multiprocessor with one processor dedicated to I/O. (The
rumour was that for the 750, the CI adapter was actually the faster CPU...)

So this is quite an old trick, and not really that difficult to do.  I would
assume that current Unix/workstation systems do something similar.

	Jan