Andy Green wrote: > So just to recap -- by not shooting ourselves in the foot with the > broken image generation, we end up with the performance anyone else > would have started with...
Yes, that one was a biggie. > and HW ECC started working because we moved > to 2.6.24, again anyone else would just tick that box now. There were two obstacles: one, the correction algorithm was broken. For one single-bit error, you got two after "correction". And if it found that the error was uncorrectable (i.e., more than one bit wrong), it happily returned "everything's fine" :-( (This is, by the way, also what every GTA01 is doing right now. So this looks like a worthy fix for the 2.6.22 tree as well. I'll add it in a little while.) Two, under the assumption that reads were for 2kB blocks, the hardware ECC was disabled, because the way the hardware was used, it would only cover 512 bytes. Turns out that reads are in fact 512 bytes, so the ECC is fine. But yes, all that are correctness considerations, not strictly performance. Still, it's kinda nice if the ECC doesn't actually make things worse ;-) > Still, from where we were yesterday, it is a big leap. Yeah, the summaries also helped a lot in the end. - Werner
