Bug 253337

+++ This bug was initially created as a clone of Bug #220902 +++

Summary:  AIO calls in the kernel do not get registered by iostat, top and
similar programs as contributing to iowait.  

Details:  When the aio kernel module reads or writes, it doesn't schedule the
I/O in a way that gets recognized by our sysstat programs.  That leads to a
situation where inaccurate outputs of iowait.  The customer's onsite GPS
consultant observed this behavior and proposed the aio.patch file that's
attached to this IT ticket.  It was submitted to rhkernel-list and lkml where it
was rejected by lkml.

-- Additional comment from jmoyer on 2007-08-17 12:25 EST --
Created an attachment (id=161753)
Account for io_wait time properly when using AIO

In theory, not all AIO operations actually initiate disk I/O.  The only AIO
command that does not issue disk I/O is IOCB_CMD_NOOP which, as it happens,
actually isn't usable as implemented!

If IOCB_CMD_NOOP were usable, the attached patch would be the right way to
account for I/O wait time appropriately.  It creates a hash table of io
contexts and associated I/O counters.  Whenever a process schedules, this code
checks to see if there is real I/O pending, and calls io_schedule if so. 
Otherwise, it uses schedule() as before.

Given that IOCB_CMD_NOOP doesn't actually work, the real solution is to just
call io_schedule unconditionally.  That patch is much simpler, and I'll attach
it after discussing this with upstream.

I'm posting this patch in case we ever wanted to correctly implement the
IOCB_CMD_NOOP command.