(An accident due to an old configuration error. I had a 200 MB swap
partition, of which only 128 MB worked. So I swapoffed it, then split
it into two 100 MB partitions. But cfdisk reported a failure rescanning
the partition table (presumably because some other partitions on that
drive were in use, like the root filesystem) and told me I'd have to
reboot to get the changes recognized. (Ugh; maybe I can fix that
windoze-ism.)
So I commented the swap entry out of /etc/fstab, since I hadn't mkswapped
the second partition, and the first still looked like 128 MB, which
might cause problems.
I was going to fix it all up next time I rebooted (run mkswap and swapon),
but never quite got around to it. (I only noticed now because I was
looking up the exact configuration for the bug report.)
Anyway, when a lot of stuff is going on, my system (Debian 2.0, 2.1.92,
i586-133, 96 MB) goes into swap death - it hits the disk for varying
lengths of time every second on the second, and while it appears to
still be crawling along, it takes 5 minutes to do anything.
I lose patience trying to kill processes and just power it down.
(Haven't had the slightest fsck problems despite this happening
about once a day since 2.1.91.)
Nothing in the syslog, and from X I can't see console messages.
This happens when I push the machine, e.g. running Netscape (I haven't
got Mozilla compiled yet - 2.1 and glibc needs some work. :-(), CVS from
vger, and dselect's install phase in this latest instance. In the past,
it's been kernel compiles that pushed it over the edge.
Anyway, I just put my swap space back and I'll see what happens now.
Yeah, the mm system needs more overhauling. We need a page-relocate
operation, so when we need contiguous free space, we can move
a physical page without affecting its virtual mappings.
Sort of a swap-out-then-in operation without the disk I/O.
Then drivers looking for contiguous memory, or memory below 16M,
could just find a likely range, and evict whoever's using the space.
Anyway...
-- -Colin- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu