Messages in this thread | | | Date | Tue, 21 Dec 1999 02:47:04 -0800 (PST) | From | Jeremy Fitzhardinge <> | Subject | Re: readX/writeX semantic and ordering |
| |
On 20-Dec-99 Gabriel Paubert wrote: > On Mon, 20 Dec 1999, Gerard Roudier wrote: >> I have been reported some problem on the G3 that seems to disappear when >> the cache is set write-through. Could a bridge misconfiguration explain >> that ? > Do you have any pattern about which type of G3 (uP revision, size of > backside L2 cache, output of lspci, especially host bridge) ?
MPC750 ("Arthur") rev 2.2, 1MB backside (also happens with 512k), bridge is MPC106 ("Grackle"), rev 4.0.
With L2 in copy-back, it appears the sym'896 mis-DMAs stuff from system memory, and/or gets out of sync with the IO operations into PCI space. With L2 in write-through mode, it appears to work OK, even though L1 is still copy-back and enabled. When it fails, it fails after about .5-1 hour of heavy SCSI activity; I've never observed a failure with L2 in write-through, but without a clear understanding of what the failure it, I'm not convined its fixed (as opposed to merely much less likely).
The '106 has very few configuration options which could affect snooping, and none apply to single CPU systems (I think). The '106 always sets /GBL, which should cause the L2 to be snooped (there is a 750 bug where the L2 could be mis-snooped with non-GBL 60x bus cycles).
J
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/
| |