On Sun, Jun 21, 2009 at 02:39:31AM +0200, Carl-Daniel Hailfinger wrote:
On 20.06.2009 20:08, Ward Vandewege wrote:
[...] but I'll still have APs and BSP talking all at once, which I'm also seeing on K10.
So mangled printk problem is visible on K8 and Fam10?
Yes. Here's a K8 boot that works reliably on h8dme (r4314) but has some garbled output. Check right after it says "Clearing initial memory region: Done".
http://ward.vandewege.net/coreboot/h8dme/minicom-20090618q.cap
And here's a K10 boot that does not boot because there's something going on with the latest microcode for this cpu (but that's another issue I think). It has the garbled output right after "Start node 01 done." and eventually soft resets:
http://ward.vandewege.net/coreboot/h8dmr/fam10/h8dmr-am.cap
I just read through the M57SLI early code path again and it seems we really have a race condition which causes stack corruption on SMP, possibly only visible if there is more than one AP. Or caches of each core are independent during early startup and the "every AP can use RAM" line of thought needs to be revised.
Interesting. The K8 machine above is a supermicro h8dme with 2x dual core, and the K10 machine is a supermicro h8dmr with 2x quad core. The m57sli only has one cpu socket of course but I use it regularly with dual core cpus, and I have never noticed garbled output. I could be overlooking it of course, it's relatively subtle in the K8 example.
Thanks, Ward.