Patch is also available here:
http://www.theshore.net/~caker/patches/

Apply to a tree already patched with 2.4-um or 2.6-um.

-Chris

---token-limiter-v1.README---

This patch implements a simple token bucket filter/limiter on requests
processed through the async io_thread.  It will work with or without COW,
but not on ubd devices in sync mode.

* Two new command line arguments for setting at boot-time:

token_max=60000
    This sets the maximum tokens in the token-bucket I/O request limiter.
    Each io_thread request subtracts one token from the bucket.  The bucket
    is supplied with <token_refill> tokens every second.  When the bucket
    becomes empty, I/O requests are throttled to the <token_refill> rate.

token_refill=2000
    This sets the bucket's refill rate, adding <token_refill> tokens to the
    bucket every second. This becomes the I/O request rate when the bucket
    becomes empty.

* Three new mconsole commands to get status information and change the two 
variables above:

(mconsole) help
    io_status - Return current I/O status and settings
    io_token_max <num> - sets the bucket size
    io_token_refill <num> - number of tokens to add each second

* Example io_status output while:

idle:
io_count=1845202 io_rate=0 io_tokens=50000 token_refill=1024 \
token_max=50000

bursting:
io_count=1861594 io_rate=6806 io_tokens=35754 token_refill=1024 \
token_max=50000

throttled:
io_count=1994811 io_rate=1087 io_tokens=-153 token_refill=1024 \
token_max=50000 
io_count=2052289 io_rate=1057 io_tokens=-286 token_refill=1024 \
token_max=50000
io_count=2067794 io_rate=1015 io_tokens=-132 token_refill=1024 \
token_max=50000

Our sample shows io_tokens being negative, but - while waiting for the bucket 
to be re-supplied, requests will still dribble through no faster than 
token_refill req/second. Or, perhaps this is due to inaccuracies of nanosleep
or a bug in my logic.

* TODO
Stop using an alarm in the io_thread, and instead use gettimeofday to 
calculate the token credits due and io_rate.