Re: Any news on a higher performance sata_sil SIL_QUIRK_MOD15WRITEworkaround?

From: Jeff Garzik
Date: Thu Aug 12 2004 - 01:44:48 EST


Clem Taylor wrote:
I've been really disappointed by the performance of the Silicon Image 3114 on my new x86_box. I spent a bunch of time looking into the problem, thinking it was a software RAID5 or xfs issue causing 4K IOs.
I don't know why I didn't notice the 'applying Seagate errata fix' in dmesg until after I did a bunch of performance testing and realized that it was a sata_sil issue.

So, I was wondering what I can do about this problem? I am not currently

Get a different controller + disk combination.


getting enough disk performance to justify the amount spent on the system or enough to satisfy the application I'm working on. Before I go out and purchase a 3ware controller and re-install the machine (ouch), is there any chance of a better work around in the near future? I'd be more than willing to test out a patch.

Is the problem with really with nblocks % 15 == 1? Or is the problem with nblocks % 15 == 0? If it is the later and I'm using xfs with 4K blocks, couldn't I just turn off the workaround or will the RAID5 driver potentially break up larger requests?

The problem is that the Silicon Image controller sends unusual -- but legal -- block sizes to the SATA device.

Older Seagates cannot cope with this unique, but spec-correct behavior.

This issue cannot even be worked around with "nblocks % 15 == 1", as was previously thought. Using that formula just makes the problem harder to reproduce.

Further, I don't have any plans to address the performance issue, since the set of affected drives is finite.


It would seem that the root of the problem is a Seagate issue. Does anyone know if Seagate fixed the problem with a firmware update?

You could find out for us, and let us know :)

Jeff


-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/