Hello,
I wrote some month ago about a problem with last kernel release. The
thread is accessible here:
http://marc.theaimsgroup.com/?l=linux-kernel&m=103875989402489&w=2
To sum-up the problem: I have a tyan tiger MPX with AMD762 and AMD768
chipsets. With latest kernel releases, every time I listen music and
that something appens on my screen (moving a window, watching a movie)
the sound stutter. The problem occurs whatever the soundcard and the
graphic card is. But the problem didn't occured with older kernel
releases like 2.4.8. I received tons of advices, but none was working.
Now I have done some investigations, and I think I have found the
problem: It has appeared between kernel-2.4.18-pre9 and kernel-2.4.18-rc1
If I am not wrong, between these two versions, Alan Cox did a change in
drivers/pci/quirks.c and this is this change which cause the problem.
*** Before the change we got:
static void __init quirk_amd_ordering(struct pci_dev *dev)
{
u32 pcic;
pci_read_config_dword(dev, 0x42, &pcic);
if((pcic&2)==0)
{
pcic |= 2;
printk(KERN_WARNING "BIOS disabled PCI ordering
compliance, so we enabled it again.\n");
pci_write_config_dword(dev, 0x42, pcic);
}
}
*** After the change we got:
/*
* Following the PCI ordering rules is optional on the AMD762. I'm not
* sure what the designers were smoking but let's not inhale...
*
* To be fair to AMD, it follows the spec by default, its BIOS people
* who turn it off!
*/
static void __init quirk_amd_ordering(struct pci_dev *dev)
{
u32 pcic;
pci_read_config_dword(dev, 0x4C, &pcic);
if((pcic&6)!=6)
{
pcic |= 6;
printk(KERN_WARNING "BIOS failed to enable PCI standards compliance,
fixing this error.\n");
pci_write_config_dword(dev, 0x4C, pcic);
pci_read_config_dword(dev, 0x84, &pcic);
pcic |= (1<<23); /* Required in this mode */
pci_write_config_dword(dev, 0x84, pcic);
}*/
}
Continuing my investigations, I have looked at the AMD762 datasheet. The
new code which is intented to follow the PCI standards causes my
problem! In fact with this code the system becomes completly slow, quite
unusable.
I like the comment for the fix, but I have an other version :-)
I think AMD didn't follow the PCI specs by default (and no it's not bios
people who turned it off, it's AMD!) because their chipsed is bogus and
didn't work well (at all) following the specs !
I may be wrong but I don't think so. What do you think?
As for the fix, I propose to remove the quirk_amd_ordering function, or
at least allow to don't use it. In fact the old kernels (before
2.4.18-rc1) worked fine for me because they didn't contained this
workaround. Old kernels contained a workaround, but it was IMHO false
because it changed reserved and probably unassigned registers (0x42).
And note that due to this problem, I am using kernel-2.4.8 for about a
year now, with 6 of the 7 PCI and AGP slots used, and I didn't noticed
any instability.
I hope I was clear, and I really would like to see this problem solved.
Regards,
Xavier IZARD
P.S. I have also tested the 2.4.22-pre4 kernel, but as it contains the
same code, the problem is still here.
--
E-mail:
[email protected] [email protected]
Please no longer use [email protected], this e-mail will be removed soon.
Homepage:
http://xizard.free.fr
http://www.chez.com/xizard/
On Gwe, 2003-07-11 at 18:02, xi wrote:
> Now I have done some investigations, and I think I have found the
> problem: It has appeared between kernel-2.4.18-pre9 and kernel-2.4.18-rc1
> If I am not wrong, between these two versions, Alan Cox did a change in
> drivers/pci/quirks.c and this is this change which cause the problem.
As I read the documentation the other change is also required in this
situation to avoid a chipset lockup. It might be worth you rechecking
the AMD errata docs for 762/768 again to be sure I didnt screw up and
there are not newer rules for other revisions.
Alan Cox wrote:
> On Gwe, 2003-07-11 at 18:02, xi wrote:
>
>>Now I have done some investigations, and I think I have found the
>>problem: It has appeared between kernel-2.4.18-pre9 and kernel-2.4.18-rc1
>>If I am not wrong, between these two versions, Alan Cox did a change in
>>drivers/pci/quirks.c and this is this change which cause the problem.
>
>
> As I read the documentation the other change is also required in this
> situation to avoid a chipset lockup. It might be worth you rechecking
> the AMD errata docs for 762/768 again to be sure I didnt screw up and
> there are not newer rules for other revisions.
>
Ok, I have rechecked the errata docs. I have not found any
recommandation change for PCI compliance configuration registers.
And one interesting thing:
in the AMD762 datasheet (24462.pdf) page 231 (Recommanded BIOS
settings), I can see this: "Numerical Values shown with h or b are
preferred settings." ; and AMD recommand this:
-> set bits 2 and 1 of register 0x4C to "0b"
-> set bits 23 and 3 respectively to "0b" and "1b"
I can confirm that these settings works much more better, even if they
don't exactly follow PCI specs. And I don't think this is specific to my
cards since I have tested others.
Furthermore, my AMD762 is revision B1 (just before the last one: C0),
and my AMD768 revision is B2, the last one.
Would you accept I make a patch which doesn't make any change in these
registers at least up to AMD762 revision B1 (ie keeping recommanded
values from AMD) ?
Or could you propose an other solution ?
Regards,
Xavier
--
E-mail:
[email protected] [email protected]
Please no longer use [email protected], this e-mail will be removed soon.
Homepage:
http://xizard.free.fr
http://www.chez.com/xizard/
On Gwe, 2003-07-11 at 22:04, xi wrote:
> And one interesting thing:
> in the AMD762 datasheet (24462.pdf) page 231 (Recommanded BIOS
> settings), I can see this: "Numerical Values shown with h or b are
> preferred settings." ; and AMD recommand this:
> -> set bits 2 and 1 of register 0x4C to "0b"
> -> set bits 23 and 3 respectively to "0b" and "1b"
>
> I can confirm that these settings works much more better, even if they
> don't exactly follow PCI specs. And I don't think this is specific to my
> cards since I have tested others.
> Furthermore, my AMD762 is revision B1 (just before the last one: C0),
> and my AMD768 revision is B2, the last one.
>
> Would you accept I make a patch which doesn't make any change in these
> registers at least up to AMD762 revision B1 (ie keeping recommanded
> values from AMD) ?
Lets try the AMD recommended settings. My old doc doesnt seem to have
those. I'll by happy to trial the patch in -ac and see if it plays up
the usual suspects for PCI spec violations (tg3 and i2o)
--- quirks_original.c 2003-07-12 01:58:17.000000000 +0000
+++ quirks.c 2003-07-12 04:44:02.000000000 +0000
@@ -464,23 +464,24 @@
* Following the PCI ordering rules is optional on the AMD762. I'm not
* sure what the designers were smoking but let's not inhale...
*
- * To be fair to AMD, it follows the spec by default, its BIOS people
- * who turn it off!
+ * In fact, AMD even recommends to don't follow PCI standards
+ * in the section "Recommended BIOS settings" of the datasheet
*/
static void __init quirk_amd_ordering(struct pci_dev *dev)
{
u32 pcic;
+
+ printk(KERN_WARNING "Setting AMD recommended values for PCI bus. It isn't fully PCI standards compliant\n");
+
pci_read_config_dword(dev, 0x4C, &pcic);
- if((pcic&6)!=6)
- {
- pcic |= 6;
- printk(KERN_WARNING "BIOS failed to enable PCI standards compliance, fixing this error.\n");
+ pcic &= ~((u32)6);
pci_write_config_dword(dev, 0x4C, pcic);
+
pci_read_config_dword(dev, 0x84, &pcic);
- pcic |= (1<<23); /* Required in this mode */
+ pcic &= ~((u32)(1<<23));
+ pcic |= (1<<3);
pci_write_config_dword(dev, 0x84, pcic);
- }
}
#ifdef CONFIG_X86_IO_APIC
On Sad, 2003-07-12 at 06:08, xi wrote:
> * patch_AMD762_PCI_compliance_set_by_bios.diff : lets the BIOS decide
> about PCI configuration
Breaks i2o with a 1004 BIOS on the ASUS dual athlon setup
>
> * patch_AMD762_PCI_default_AMD_settings.diff : follow AMD
> recommendations but not PCI specs.
Looks better