Abysmal performance for Intel MatrixRAID

[Michael Graff <explorer%flame.org@localhost>]

I have a machine here which has the Intel MatrixRAID controller in it,
and would like to use it.  I am running 5.1_RC4 amd64 on this box.

Without applying any source patches, it is unusable.  The machine locks
up.  I found this to be a solution:

http://mail-index.netbsd.org/tech-kern/2008/09/17/msg002734.html

However, with this patch applied, the disk tops out at 8.9 MB/sec write,
which is quite a bit less than the non-RAID disks on the machine.  I am
not including the read speeds as these are all from cache anyway with
this size of test file.

              -------Sequential Output--------
              -Per Char- --Block--- -Rewrite--
Machine    MB K/sec %CPU K/sec %CPU K/sec %CPU
ld0      1024  9690  5.5  8941  2.4  9677  3.3
wd2      1024 44196 25.6 39776  8.1 40056 12.6
wd3      1024 26923 15.9 25109  4.9 25179  7.7

While the disks wd2 and wd3 are not impressive either, at least they are
around what I'd accept as OK.

Just for comparison, this same drive non-raid came to be around 48
MB/sec in the block column.  I don't have a spare non-raid drive to try
at the moment.

I also found
http://www.mail-archive.com/tech-kern%netbsd.org@localhost/msg01524.html, which 
I
assume means this patch has been applied to -current and also mentioned
another patch which fixed some other issues.  I've applied both of these
to my netbsd_5 tree before running the above tests.

When doing even small operations to this drive, it is pegged at 100%
busy by even simple operations.  Part of this may simply be bandwidth,
but I suspect it is caused by the patch from msg002734 above.  This
patch defers the VOP_STRATEGY() call through a soft interrupt rather
than handling it directly, which is done apparently to avoid a locking
problem.

--Michael