🚀 go-pugleaf

Johnny Billquist  <b...@softjar.se> wrote:
+---------------
| John Levine wrote:
| > In terms of getting useful work out of a computer that ran at about
| > 5MHz (yes, I know the KA was async) and had 1MB of core, TOPS-10 wins
| > just as much.  It had features that mattered then but don't any more
| > like doing disk I/O directly in and out of buffers in user space
| > without an extra copy.  I could have done without the sequence numbers
| > with the flag in the low bit of the word, though.
| 
| Disk directly to buffers in user space in Linux? Sure, it's called mmap. 
| Which, by the way, is more or less a direct ripoff from TOPS-20.
+---------------

Actually, TOPS-10 user-space buffered I/O was closer to Linux/Unix's
DIRECT_IO mode than to mmap(), except it was naturally multi-buffered.

That is, you had a *ring* of user-space buffers, with ownership flags
so both the kernel and the user program could see who was working
on a given buffer. This quite naturally allowed kernel I/O and the
user program to work asynchronously, as long as neither side caught
up to the portion of the ring owned by the other side. [And if that
happens, the faster side simply stopped until there was a free (or full)
buffer available for it.]

Plus, you could have *multiple* devices sharing the same user-space
buffer ring, such as disk I/O and magtape I/O, allowing the data to
be streamed continuously from one device through the buffers to the
other device. [In this case the ring was paritioned into *four* areas:
device #1, user area #1, device #2, user area #2.] This is how the
BACKUP program managed to keep the magtapes writing at full speed
[if you had fast-enough disks].


-Rob

-----
Rob Warnock		<rp...@rpw3.org>
627 26th Avenue		<http://rpw3.org/>
San Mateo, CA 94403