On Wed, 2003-05-28 at 11:39, Jens Axboe wrote:

> Correction then, it doesn't appear to be starvation in the usual sense.
> But you are right, pulling some stats out of the situation would be
> nice. I still can't reproduce here.

Well, it's not pretty but it gets some numbers out there.  This patch
only calculates the time spent waiting in __get_request_wait, it isn't
interested in any other metrics.  stats are per-queue and are reset when
you mount the FS, you get a print out either when you unmount the FS or
when you run elvtune /dev/xxx (no other args, just enough to trigger the
read ioctl).

The output looks like this (after a dbench 50 run 2.4.21-rc6)

device 03:04: num_req 12248, total jiffies waited 26729
        417 forced to wait
        1 min wait, 432 max wait
        64 average wait
        314 < 100, 62 < 200, 20 < 300, 20 < 400, 1 < 500
        0 waits longer than 500 jiffies

It tells us there were 12248 total requests (merges don't count), and
that we spent 26,729 jiffies waiting in __get_request_wait.  We had to
wait 417 times, the minimum was 1 and the max was 432 jiffies.  The line
with the < signs is a simple way to get the deviations.  314 requests
waited < 100 jiffies, 62 requests waited less than 200 jiffies, etc.

People who see stalls on UP machines and have seen improvements by
playing with code in drivers/block/ll_rw_blk.c are encouraged to try
getting numbers with this patch applied.  It will make it easier to
figure things out.

I haven't tried Andrea's fix-pausing on top of this yet, any rejects
should be minor.

-chris