Migrating from 2.6.11 to 2.6.23 breaks pci-e with LSI 1068 SAS chip (solved!)

[Vince Asbridge]

Solved!

Problem summary.

LSI logic devices (1068e, 1064e, fc949e) do not work with uBoot version
1.3.0 and Linux kernel 2.6.23, where they work perfectly fine in kernel
2.6.11.

Symptom is that they do not show up at all to Linux once booted (lspci shows
nothing).

We obtained a pci-e analyzer and found the following:

uBoot V1.3.0 now scans the pci configuration, and when it does it assigns
bus numbers to the pci-e devices it finds.  These numbers that are assigned
are different from those that are assigned by Linux when the system boots.

It is legal to change pci-e bus numbers "on the fly" but doing so requires a
config write cycle to the pci-e device, because the pci-e spec states that
the device must register a new bus number if a config write cycle is issues
with the new bus number.

When we boot the Freescale machine with uBoot 1.3.0 and kernel 2.6.23, the
bus number under uBoot gets assigned to 1, and the bus number under Linux
gets assigned to 4.  Between the change of bus numbers we do not see a
config write go to the LSI device with the new bus number, and therefore it
continues to respond on bus 1 and ignore config reads to bus 4.

We temporarily fixed the issue by defining the CONFIG_PCI_NOSCAN flag under
uBoot, which causes the LSI device not to get assigned a bus number at
uBoot, and therefore when Linux begins doing config reads at bus 4 to the
LSI device registers the bus number and responds correctly.

We will do a bit more investigation to see if this has already been fixed in
a newer kernel, but for now not scanning at uBoot fixes the issue.

It's a bit of a mystery why the LSI devices behave differently from any
other pci-e device we put in the system, but they seem to be adhering to the
letter of the specification.

Vince Asbridge

------- original post -------

All,

I'm new to this mailing list, but have not had any luck finding information
on this issue.

Please be kind if I break the forum rules on my first post.

We recently tried to upgrade our Freescale CDS 8548 look-alike module (code
name ATCA1000) from the 2.6.11 based BSP to the 2.6.23 based BSP.

The upgrade went fairly smoothly, until we tried using SOME pci-e devices
(some work fine, some don't show up to lspci).

LSI pci-e controllers no longer show up at all!

We see the ixgbe (intel 10G), SiliconImage SATA controller but do not see
LSI devices (Specifically 1068 SAS, FC949-E fibrechannel).

We're guessing it's a resource issue behind the bridge, because the LSI
devices try to allocate 1 - 3M behind the bridge, but we can't find the bug,
or where we would debug such an issue.

The devices seem to "train" correctly, because we have an LED on the pci-e
switch (PLX 8 port pci-e switch), and it's ON indicating pci-e link between
the bridge and the 1068 device).

We're totally at a loss as to why this always worked on the 2.6.11 kernel
but doesn't work on 2.6.23.

Using lspci, the LSI adapters do not show up in the list at all, as though
they are not plugged into the system.

Is there something that needs to be done with respect to PCI-E devices that
is new in the 2.6.23 based BSP that did not need to be done in the 2.6.11
based kit?  For example, are pci resources allocated by a different piece of
code, that may have some issue allocating resources for the LSI adapters?

I currently do not have access to a pci-e analyzer.

Thanks for any help,
Vince Asbridge



-------------- next part --------------
A non-text attachment was scrubbed...
Name: winmail.dat
Type: application/ms-tnef
Size: 3002 bytes
Desc: not available
URL: <http://ozlabs.org/pipermail/linuxppc-embedded/attachments/20080702/b1ad39ff/attachment.bin>