Re: Bad write performance with libata-tj-stable-2.6.17.4-20060710, pcmcia based sata_sil24, PMP, and NCQ drive

[Derek Taubert <taubert@geeks.org>]

On Fri, Sep 01, 2006 at 10:34:59PM -0700, Derek Taubert wrote:
> I'm now suspecting that the performance problems are due to bus issues
> on the PCMCIA/PCI side.  Will poke around there and maybe try this with
> a different laptop...

Found the problem!!!!

The Cache Line Size field in the 3124 PCI Config Space needed to be
set properly.  I suspect that this caused the 3124 DMA engine to issue
non-burst reads for all data going toward the disk.  Since my old
laptop has two bridges between host memory and the 3124, it surely made
for some long read latencies (and thus very bad throughput).

Before:
root@linux-srv:~ # od -Ax -t x4 /sys/bus/pci/devices/0000:06:00.0/config
000000 31241095 02b00087 01800001 00004000
000010 16008004 00000000 16000004 00000000
000020 00001c01 00000000 00000000 31241095
000030 00000000 00000064 00000000 0000010a
000040 00525407 12c3fff8 00000000 00000000
000050 00000000 00800005 00000000 00000000
000060 00000000 06224001 19002000 00000000
000070 00000000 00000000 00000000 00000000

I copied /sys/bus/pci/devices/0000:06:00.0/config to a temp file, used
bvi to edit it, and then copied it back using dd:
dd if=/tmp/sil24_config of=/sys/bus/pci/devices/0000:06:00.0/config bs=4 skip=3 seek=3 count=1 conv=notrunc

After:
root@linux-srv:~ # od -Ax -t x4 /sys/bus/pci/devices/0000:06:00.0/config
000000 31241095 02b00087 01800001 00004008
000010 16008004 00000000 16000004 00000000
000020 00001c01 00000000 00000000 31241095
000030 00000000 00000064 00000000 0000010a
000040 00525407 12c3fff8 00000000 00000000
000050 00000000 00800005 00000000 00000000
000060 00000000 06224001 19002000 00000000
000070 00000000 00000000 00000000 00000000

root@linux-srv:~ # dd if=/dev/zero of=/dev/sda1 bs=1M count=1000
1000+0 records in
1000+0 records out
1048576000 bytes (1.0 GB) copied, 16.7798 seconds, 62.5 MB/s

Phew!  I was just about to spend a ridiculous amount of money on ebay for
a logic analyzer...

I'm not sure if this will affect PCI-X configurations, but it is
certainly worth trying.

Derek

-- 
VGER BF report: H 0.104804