pata_ali can't detect partitions
Hello,
yesterday I tried to install archlinux-i586 on a old computer I have. It's a K6-3 with a ALI chipset.
I noticed that the system was able to find the disk as sda but could not recognize the partitions on the drive. This happens with the 2.6.30-ARCH kernel of the archlinux-i586 installer.
I noticed also that with an older kernel (2.6.28 IIRC) the system could detect all the partitions.
So yesterday I had some spare time to try to bisect the problem with git bisect. The result of the bisection was the following commit:
commit 1b2c357c301b76118973763e886d9f70a7f50f7epata_ali: force initialise a few bits
I tried to backout the commit but that didn't succeed. So I looked carefully at the diff and I came out with the attached patch which seems to fix the problem (at least now the kernel detects the partitions correctly).
Since I saw that the above commit is fixing some real bugs (sunblade related) I don't know whether the attached patch is right or is backing out the intended fix. All I know is that it works for me and for this reason I'm reporting it.
Thanks,
Gianluca
drivers/ata/pata_ali.c renamed from drivers/ata/pata_ali.c.orig +1 -1
@@ -453,7 +453,7 @@/* Clear CD-ROM DMA write bit */ tmp &=0x7F ;/* Cable and UDMA */ -pci_write_config_byte (pdev,0x4B , tmp |0x09 ); +pci_write_config_byte (pdev,0x4B , tmp |0x08 );/* * CD_ROM DMA on (0x53 bit 0). Enable this even if we want * to use PIO. 0x53 bit 1 (rev 20 only) - enable FIFO control
Quoted 6 lines
commit
1b2c357c301b76118973763e886d9f70a7f50f7epata_ali: force initialise a few bitsI tried to backout the commit but that didn't succeed. So I looked carefully at the diff and I came out with the attached patch which seems to fix the problem (at least now the kernel detects the partitions correctly).
Thanks for doing this. Can you send me an lspci -vvxxx and a dmesg of the machine in question.
Quoted 4 lines
Since I saw that the above commit is fixing some real bugs (sunblade related) I don't know whether the attached patch is right or is backing out the intended fix. All I know is that it works for me and for this reason I'm reporting it.
Some of the registers involved are deep magic (even with the official documentation its unclear what should occur in all cases). I will have another dig - probably it depends on the chip rev what the right setting is.
Alan
As requested I'm attaching the output of the commands lspci -vvxxx and dmesg. Since they were quite big I gzipped them.
You can find the uncompressed copy also here for your convenience:
http://www.sottospazio.it/lspci-k63d
http://www.sottospazio.it/dmesg-k63d
Thanks,
Gianluca
From: gianluca <[email protected]>
Date: Tue, 10 Nov 2009 21:25:33 +0100
Quoted 7 lines
As requested I'm attaching the output of the commands lspci -vvxxx and dmesg. Since they were quite big I gzipped them.
You can find the uncompressed copy also here for your convenience:
http://www.sottospazio.it/lspci-k63d
http://www.sottospazio.it/dmesg-k63d
I want to help with this so I'm going to fire up some of my ALI using sparc64 systems and try to give a hand.
From: Alan Cox <[email protected]>
Date: Tue, 10 Nov 2009 11:11:25 +0000
Quoted 4 lines
Some of the registers involved are deep magic (even with the official documentation its unclear what should occur in all cases). I will have another dig - probably it depends on the chip rev what the right setting is.
So while getting my sunblade up and going, I took a look at the patch in question. The northbridge logic changed a bit and I suspect this is part of the problem.
In the IDE layer driver, the guard is:
if (north && north->vendor != PCI_VENDOR_ID_AL) goto out;
This means the programming is done iff:
1) We find no device at PCI_DEVFN(0,0)
2) We find a device and vendor is ALI
I suspect case #1 triggers on sparc64. The guard in the PATA driver is:
north = pci_get_bus_and_slot(0, PCI_DEVFN(0,0));
if (north && north->vendor == PCI_VENDOR_ID_AL && ali_isa_bridge) {which is different. It won't do the programming if we find no device at PCI_DEVFN(0,0).
This might be the critical difference, I don't know, just pointing it out.
I'll do some checks once my slow test system is up and going.
Quoted 14 lines
In the IDE layer driver, the guard is:
if (north && north->vendor != PCI_VENDOR_ID_AL) goto out;This means the programming is done iff:
1) We find no device at
PCI_DEVFN(0,0)2) We find a device and vendor is ALI
I suspect case #1 triggers on sparc64. The guard in the PATA driver is:
Quoted 8 lines
north = pci_get_bus_and_slot(0, PCI_DEVFN(0,0)); if (north && north->vendor == PCI_VENDOR_ID_AL && ali_isa_bridge) {which is different. It won't do the programming if we find no device at
PCI_DEVFN(0,0).This might be the critical difference, I don't know, just pointing it out.
The logic in pata_ali is the intended logic there (and the comments in drivers/ide match the code in pata_ali but the code doesn't)
What docs I have say you need to program the isa enable bit if you are using an ALi north bridge (which always appears at 0,0,0). The other PC case is the Transmeta systems, and in that case the setup is done by the firmware and undocumented entirely.
I have exactly zero Sparc64 documentation on the subject. The only Sparc related material I have is the C2/C3 post reset stuff. That *might* be a relevant difference in behaviour for C2 and C3 devices but only with an ALi bridge (ali_c2_c3_postreset)
Alan
From: Alan Cox <[email protected]>
Date: Thu, 19 Nov 2009 21:10:31 +0000
Quoted 4 lines
I have exactly zero Sparc64 documentation on the subject. The only Sparc related material I have is the C2/C3 post reset stuff. That *might* be a relevant difference in behaviour for C2 and C3 devices but only with an ALi bridge (
ali_c2_c3_postreset)
So I can't even reproduce these ALI problems on my sb100, it works perfectly fine with Linus's tree. I'll play around on my sb1000 and some other drives.
And yes that poking in the ISA bridge during reset is necessary to make sure all the gates are in the proper initial state.