Skip to contentSkip to conversation

2.6.14-rc1 breaks tg3 on ia64

DISCUSSION Keith Owens 6 participants 16 messages Sep 13, 2005 – Sep 18, 2005

lore
Share & tools
lore
#0 lore raw

2.6.14-rc1 + kdb on ia64 (SGI Altix).

tg3.c:v3.39 (September 5, 2005)
ACPI: PCI Interrupt 0001:01:04.0[A]: no GSI
BRIDGE ERR_STATUS 0x800
BRIDGE ERR_STATUS 0x800
PCI BRIDGE ERROR: int_status is 0x800 for 011c32:slab0:widget15:bus0
Dumping relevant 011c32:slab0:widget15:bus0 registers for each bit set...
11: PCI bus device select timeout
PCI Error Address Register: 0x3000000316808
PCI Error Address: 0x316808
PIC Multiple Interrupt Register is 0x800
11: PCI bus device select timeout

Followed by a machine check and reboot :( 2.6.13 worked fine. Any ideas which patch to backout this time?

#1 lore raw

From: Keith Owens <[email protected]>
Date: Tue, 13 Sep 2005 15:22:17 +1000

Quoted 14 lines

2.6.14-rc1 + kdb on ia64 (SGI Altix).

tg3.c:v3.39 (September 5, 2005)
ACPI: PCI Interrupt 0001:01:04.0[A]: no GSI
BRIDGE ERR_STATUS 0x800
BRIDGE ERR_STATUS 0x800
PCI BRIDGE ERROR: int_status is 0x800 for 011c32:slab0:widget15:bus0
Dumping relevant 011c32:slab0:widget15:bus0 registers for each bit set...
11: PCI bus device select timeout
PCI Error Address Register: 0x3000000316808
PCI Error Address: 0x316808
PIC Multiple Interrupt Register is 0x800
11: PCI bus device select timeout

Quoted 2 lines

Followed by a machine check and reboot :( 2.6.13 worked fine. Any ideas which patch to backout this time?

Does copying over the 2.6.13 tg3.[ch] driver over into your 2.6.14-rc1 tree make it work?

#2 lore raw
Quoted 17 lines · David S. Miller

From: Keith Owens <[email protected]>
Date: Tue, 13 Sep 2005 15:22:17 +1000

Quoted 14 lines

2.6.14-rc1 + kdb on ia64 (SGI Altix).

tg3.c:v3.39 (September 5, 2005)
ACPI: PCI Interrupt 0001:01:04.0[A]: no GSI
BRIDGE ERR_STATUS 0x800
BRIDGE ERR_STATUS 0x800
PCI BRIDGE ERROR: int_status is 0x800 for 011c32:slab0:widget15:bus0
Dumping relevant 011c32:slab0:widget15:bus0 registers for each bit set...
11: PCI bus device select timeout
PCI Error Address Register: 0x3000000316808
PCI Error Address: 0x316808
PIC Multiple Interrupt Register is 0x800
11: PCI bus device select timeout

Quoted 5 lines
Quoted 2 lines

Followed by a machine check and reboot :( 2.6.13 worked fine. Any ideas which patch to backout this time?

Does copying over the 2.6.13 tg3.[ch] driver over into your 2.6.14-rc1 tree make it work?

No, the 2.6.13 driver in 2.6.14-rc1 has exactly the same problem.

The last time that tg3 broke like this, it was because of the patch below, in 2.6.13-rc6. That was backed out in 2.6.13-rc7. Was the PCI patch (or equivalent) reinstated in 2.6.14-rc1?

From: John W. Linville <[email protected]>
Date: Fri, 5 Aug 2005 01:06:10 +0000 (-0700)
Subject: [PATCH] PCI: restore BAR values after D3hot->D0 for devices that need it
X-Git-Tag: v2.6.13-rc6
X-Git-Url: http://kernel.org/git/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;hþc59a711eef002d4ef9eb8de09dd0a26986eb77

#3 lore raw

From: Keith Owens <[email protected]>
Date: Tue, 13 Sep 2005 16:17:33 +1000

Quoted 5 lines

No, the 2.6.13 driver in 2.6.14-rc1 has exactly the same problem.

The last time that tg3 broke like this, it was because of the patch below, in 2.6.13-rc6. That was backed out in 2.6.13-rc7. Was the PCI patch (or equivalent) reinstated in 2.6.14-rc1?

It was reinstated, with a fix for the sparc64 problems it caused. I wasn't aware of any ia64 regressions introduced by it.

#4 lore raw
In reply to Keith Owens
Quoted 31 lines · Keith Owens
Quoted 17 lines · David S. Miller

From: Keith Owens <[email protected]>
Date: Tue, 13 Sep 2005 15:22:17 +1000

Quoted 14 lines

2.6.14-rc1 + kdb on ia64 (SGI Altix).

tg3.c:v3.39 (September 5, 2005)
ACPI: PCI Interrupt 0001:01:04.0[A]: no GSI
BRIDGE ERR_STATUS 0x800
BRIDGE ERR_STATUS 0x800
PCI BRIDGE ERROR: int_status is 0x800 for 011c32:slab0:widget15:bus0
Dumping relevant 011c32:slab0:widget15:bus0 registers for each bit set...
11: PCI bus device select timeout
PCI Error Address Register: 0x3000000316808
PCI Error Address: 0x316808
PIC Multiple Interrupt Register is 0x800
11: PCI bus device select timeout

Quoted 5 lines
Quoted 2 lines

Followed by a machine check and reboot :( 2.6.13 worked fine. Any ideas which patch to backout this time?

Does copying over the 2.6.13 tg3.[ch] driver over into your 2.6.14-rc1 tree make it work?

No, the 2.6.13 driver in 2.6.14-rc1 has exactly the same problem.

The last time that tg3 broke like this, it was because of the patch below, in 2.6.13-rc6. That was backed out in 2.6.13-rc7. Was the PCI patch (or equivalent) reinstated in 2.6.14-rc1?

Quoted 5 lines

From: John W. Linville <[email protected]>
Date: Fri, 5 Aug 2005 01:06:10 +0000 (-0700)
Subject: [PATCH] PCI: restore BAR values after D3hot->D0 for devices that need it
X-Git-Tag: v2.6.13-rc6
X-Git-Url: http://kernel.org/git/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;hþc59a711eef002d4ef9eb8de09dd0a26986eb77

Another data point. Omit the tg3 from the build and then it breaks on the SCSI qla2300 driver. It is a PCI/ACPI specific problem on ia64, not just network.

qla1280: QLA12160 found on PCI bus 1, dev 3
ACPI: PCI Interrupt 0001:01:03.0[A]: no GSI
BRIDGE ERR_STATUS 0x800
BRIDGE ERR_STATUS 0x800
PCI BRIDGE ERROR: int_status is 0x800 for 011c32:slab0:widget15:bus0
Dumping relevant 011c32:slab0:widget15:bus0 registers for each bit set...
11: PCI bus device select timeout
PCI Error Address Register: 0x3000000302008
PCI Error Address: 0x302008
PIC Multiple Interrupt Register is 0x800
11: PCI bus device select timeout

#5 lore raw
In reply to Keith Owens
Quoted 31 lines · Keith Owens
Quoted 17 lines · David S. Miller

From: Keith Owens <[email protected]>
Date: Tue, 13 Sep 2005 15:22:17 +1000

Quoted 14 lines

2.6.14-rc1 + kdb on ia64 (SGI Altix).

tg3.c:v3.39 (September 5, 2005)
ACPI: PCI Interrupt 0001:01:04.0[A]: no GSI
BRIDGE ERR_STATUS 0x800
BRIDGE ERR_STATUS 0x800
PCI BRIDGE ERROR: int_status is 0x800 for 011c32:slab0:widget15:bus0
Dumping relevant 011c32:slab0:widget15:bus0 registers for each bit set...
11: PCI bus device select timeout
PCI Error Address Register: 0x3000000316808
PCI Error Address: 0x316808
PIC Multiple Interrupt Register is 0x800
11: PCI bus device select timeout

Quoted 5 lines
Quoted 2 lines

Followed by a machine check and reboot :( 2.6.13 worked fine. Any ideas which patch to backout this time?

Does copying over the 2.6.13 tg3.[ch] driver over into your 2.6.14-rc1 tree make it work?

No, the 2.6.13 driver in 2.6.14-rc1 has exactly the same problem.

The last time that tg3 broke like this, it was because of the patch below, in 2.6.13-rc6. That was backed out in 2.6.13-rc7. Was the PCI patch (or equivalent) reinstated in 2.6.14-rc1?

Quoted 5 lines

From: John W. Linville <[email protected]>
Date: Fri, 5 Aug 2005 01:06:10 +0000 (-0700)
Subject: [PATCH] PCI: restore BAR values after D3hot->D0 for devices that need it
X-Git-Tag: v2.6.13-rc6
X-Git-Url: http://kernel.org/git/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;hþc59a711eef002d4ef9eb8de09dd0a26986eb77

So does reverting this patch solve the problem?

thanks,

greg k-h

#6 lore raw
Quoted 5 lines · Greg KH
Quoted 4 lines · Keith Owens

The last time that tg3 broke like this, it was because of the patch below, in 2.6.13-rc6. That was backed out in 2.6.13-rc7. Was the PCI patch (or equivalent) reinstated in 2.6.14-rc1?

Quoted 7 lines
Quoted 5 lines

From: John W. Linville <[email protected]>
Date: Fri, 5 Aug 2005 01:06:10 +0000 (-0700)
Subject: [PATCH] PCI: restore BAR values after D3hot->D0 for devices that need it
X-Git-Tag: v2.6.13-rc6
X-Git-Url: http://kernel.org/git/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;hþc59a711eef002d4ef9eb8de09dd0a26986eb77

So does reverting this patch solve the problem?

I reversing
http://kernel.org/git/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;h4b53dbcc977dbf2753a67c2b8fc1c061d74f21,
which appears to be the latest version of this patch. There was a patch reject in sparc64, but the common code was reverted. IA64 (SGI Altix) with that patch reverted now boots 2.6.14-rc1.

#7 lore raw
Quoted 7 lines
Quoted 1 line

So does reverting this patch solve the problem?

I reversing
http://kernel.org/git/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;h4b53dbcc977dbf2753a67c2b8fc1c061d74f21,
which appears to be the latest version of this patch. There was a patch reject in sparc64, but the common code was reverted. IA64 (SGI Altix) with that patch reverted now boots 2.6.14-rc1.

Anyone know anything more about this problem? I'm not seeing it on any of my systems ... so perhaps it only affects cards with a PCI bridge on them, or cards that haven't already been initialized by EFI.

-Tony

#8 lore raw
Quoted 8 lines · Tony Luck
Quoted 7 lines
Quoted 1 line

So does reverting this patch solve the problem?

I reversing
http://kernel.org/git/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;h4b53dbcc977dbf2753a67c2b8fc1c061d74f21,
which appears to be the latest version of this patch. There was a patch reject in sparc64, but the common code was reverted. IA64 (SGI Altix) with that patch reverted now boots 2.6.14-rc1.

Quoted 4 lines

Anyone know anything more about this problem? I'm not seeing it on any of my systems ... so perhaps it only affects cards with a PCI bridge on them, or cards that haven't already been initialized by EFI.

I posted a patch on Wednesday:

	http://www.ussg.iu.edu/hypermail/linux/kernel/0509.1/2193.html

The original reporter (Keith Owens <[email protected]>) confirmed this patch to fix the problem.

Thanks,

John

Signature · 2 lines
John W. Linville
[email protected]
#9 lore raw
Quoted 14 lines · John W. Linville
Quoted 8 lines · Tony Luck
Quoted 7 lines
Quoted 1 line

So does reverting this patch solve the problem?

I reversing
http://kernel.org/git/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;h4b53dbcc977dbf2753a67c2b8fc1c061d74f21,
which appears to be the latest version of this patch. There was a patch reject in sparc64, but the common code was reverted. IA64 (SGI Altix) with that patch reverted now boots 2.6.14-rc1.

Quoted 4 lines

Anyone know anything more about this problem? I'm not seeing it on any of my systems ... so perhaps it only affects cards with a PCI bridge on them, or cards that haven't already been initialized by EFI.

Quoted 6 lines

I posted a patch on Wednesday:

	http://www.ussg.iu.edu/hypermail/linux/kernel/0509.1/2193.html

The original reporter (Keith Owens <[email protected]>) confirmed this patch to fix the problem.

Yes, and a number of people objected to that patch. Care to respond to them?

thanks,

greg k-h

#10 lore raw
Quoted 17 lines · Greg KH
Quoted 16 lines · John W. Linville
Quoted 8 lines · Tony Luck
Quoted 7 lines
Quoted 1 line

So does reverting this patch solve the problem?

I reversing
http://kernel.org/git/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;h4b53dbcc977dbf2753a67c2b8fc1c061d74f21,
which appears to be the latest version of this patch. There was a patch reject in sparc64, but the common code was reverted. IA64 (SGI Altix) with that patch reverted now boots 2.6.14-rc1.

Quoted 4 lines

Anyone know anything more about this problem? I'm not seeing it on any of my systems ... so perhaps it only affects cards with a PCI bridge on them, or cards that haven't already been initialized by EFI.

I posted a patch on Wednesday:

Quoted 7 lines
Quoted 4 lines
	http://www.ussg.iu.edu/hypermail/linux/kernel/0509.1/2193.html

The original reporter (Keith Owens <[email protected]>) confirmed this patch to fix the problem.

Yes, and a number of people objected to that patch. Care to respond to them?

We are working on an SN-only workaround. No guarantee, but the person working on it is optimistic that we can fix the problem in SN code w/o making any generic changes. I should know more on Monday.

Long term, we are making SN ACPI compliant - or at leeast a lot closer.

Quoted 8 lines

thanks,

greg k-h
-
To unsubscribe from this list: send the line "unsubscribe linux-ia64" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

Signature · 4 lines
Thanks

Jack Steiner ([email protected])          651-683-5302
Principal Engineer                      SGI - Silicon Graphics, Inc.
#11 lore raw
Quoted 5 lines · Jack Steiner

We are working on an SN-only workaround. No guarantee, but the person working on it is optimistic that we can fix the problem in SN code w/o making any generic changes. I should know more on Monday.

Long term, we are making SN ACPI compliant - or at leeast a lot closer.

Well, there you go... :-)

John

Signature · 2 lines
John W. Linville
[email protected]
#12 lore raw
In reply to Greg KH
Quoted 10 lines · Greg KH
Quoted 9 lines · John W. Linville
Quoted 4 lines · Tony Luck

Anyone know anything more about this problem? I'm not seeing it on any of my systems ... so perhaps it only affects cards with a PCI bridge on them, or cards that haven't already been initialized by EFI.

I posted a patch on Wednesday:

Quoted 7 lines
Quoted 4 lines
	http://www.ussg.iu.edu/hypermail/linux/kernel/0509.1/2193.html

The original reporter (Keith Owens <[email protected]>) confirmed this patch to fix the problem.

Yes, and a number of people objected to that patch. Care to respond to them?

Care to check your email? :-)

	http://www.ussg.iu.edu/hypermail/linux/kernel/0509.1/2267.html

Basically, the concerns raised are non-issues. The new patch
merely limits the BAR restoration to those situations where it is truly needed. Anything broken in that situation was broken before the original patch as well.

The only other point was that this particular ia64 hardware is broken if you can't rewrite its BARs. That may well be, but we can accomodate it w/o losing the intent of the original patch.

Tony's post indicates that this is not a generic ia64 problem. If you want a patch along the lines of what Dave Miller and Ivan Kokshaysky advocate (i.e. something in the pci access routines for this box), then Keith or someone else w/ knowledge of (and access to) this box will need to step forward w/ a patch or at least some information.

John

Signature · 2 lines
John W. Linville
[email protected]
#13 lore raw
In reply to John W. Linville

From: "John W. Linville" <[email protected]>
Date: Sat, 17 Sep 2005 11:59:14 -0400

Quoted 6 lines

I posted a patch on Wednesday:

	http://www.ussg.iu.edu/hypermail/linux/kernel/0509.1/2193.html

The original reporter (Keith Owens <[email protected]>) confirmed this patch to fix the problem.

It fixes the problem, but it's a hack, and I, perhaps like Tony, personally would like to know why the these IA64 systems break for such a simple operation such as writing some base registers with values we've probed already.

#14 lore raw
Quoted 4 lines

It fixes the problem, but it's a hack, and I, perhaps like Tony, personally would like to know why the these IA64 systems break for such a simple operation such as writing some base registers with values we've probed already.

It does sound odd[1] that you can't rewrite the BARs to the values that they should already have. But I'm willing to wait to see what SGI's solution that fixes this inside arch/ia64/sn/ looks like before making any judgements.

-Tony

[1] odd = PCI spec violating?

#15 lore raw
In reply to David S. Miller
Quoted 7 lines · David S. Miller

From: "John W. Linville" <[email protected]>
Date: Sat, 17 Sep 2005 11:59:14 -0400

Quoted 4 lines

I posted a patch on Wednesday:

	http://www.ussg.iu.edu/hypermail/linux/kernel/0509.1/2193.html
Quoted 7 lines
Quoted 2 lines

The original reporter (Keith Owens <[email protected]>) confirmed this patch to fix the problem.

It fixes the problem, but it's a hack, and I, perhaps like Tony, personally would like to know why the these IA64 systems break for such a simple operation such as writing some base registers with values we've probed already.

Here is the mail from Mike Habeck (sgi) - he understands the problem muck better than I do:

(from Mike....)
The problem is we (sgi) don't support the ACPI pci_window stuff that is setup via ACPI (see: add_window() code). And as a result when pci_restore_bars() is called to restore the BARs, instead of the BARs getting "restored" they get wiped out. (pci_restore_bars() calls pci_update_resource() which calls pcibios_resource_to_bus()... it is that routine that is expecting the pci_window stuff being set up from the ACPI path... I think... (I don't know much about how the ACPI stuff works, John or Aaron could probably prove me right or wrong)... So John (or Aaron) is it ACPI that is suppose to setup this pci_window stuff?

I still question why this code path is taken... I don't know anything about the PCI Power Management stuff, but we shouldn't be in any power state that results in us needing our BARs restored. But I guess that really isn't the issue since sooner or later something else will end up using this pci_window stuff and we'll get burned them.

I suppose for a quick fix (to workaround this power management patch) could be to set the PCI_PM_CTRL_NO_SOFT_RESET in the cards PM capability down in PROM thus bypassing this "need_restore" code.

I suppose for a quick fix (to workaround this power management patch) could be to set the PCI_PM_CTRL_NO_SOFT_RESET in the cards PM capability down in PROM thus bypassing this "need_restore" code.

        /* If we're in D3, force entire word to 0.
         * This doesn't affect PME_Status, disables PME_En, and
         * sets PowerState to 0.
         */
        if (dev->current_state >= PCI_D3hot) {
                if (!(pmcsr & PCI_PM_CTRL_NO_SOFT_RESET))
                        need_restore = 1;
                pmcsr = 0;
        } else {
                pmcsr &= ~PCI_PM_CTRL_STATE_MASK;
                pmcsr |= state;
        }

Or in the kernel sgi device fixup code change the current_state to D0? It looks like it get's init'd to PCI_UNKNOWN (which is > PCI_D3hot) I don't know... will investigate more tomorrow

-mike

Signature · 4 lines
Thanks

Jack Steiner ([email protected])          651-683-5302
Principal Engineer                      SGI - Silicon Graphics, Inc.