LS1043A : "synchronous abort" at boot due to PCI config read

[Gilles Buloz]

Le 27/04/2018 18:56, Bjorn Helgaas a écrit :
> On Fri, Apr 27, 2018 at 12:29:32PM +0000, Gilles Buloz wrote:
>> Le 27/04/2018 10:43, Ard Biesheuvel a écrit :
>>> (add Bjorn and linux-pci)
>>>
>>> On 13 April 2018 at 19:32, Gilles Buloz <Gilles.Buloz at kontron.com> wrote:
>>>> Dear developers,
>>>>
>>>> I currently have two functional workarounds for this issue but
>>>> would like to know which one you would recommend, if any :-) I'm
>>>> using a LS1043A CPU (NXP QorIQ Layerscape) and get a "synchronous
>>>> external abort" when booting because of a PCI config read during
>>>> PCI scan.
>>>>
>>>> I'm using a custom hardware (based on LS1043ARDB) having a
>>>> PEX8112 PCIe-to-PCI bridge connected to the LS1043A to have a PCI
>>>> slot for legacy devices. This bridge only supports PCI-Compatible
>>>> config accesses (offset 0x00-0xFF).
> I would guess the PEX8112 itself has 4K of config space, but it only
> forwards 256 bytes of config space to the conventional PCI secondary
> bus.
>
>>>> On this PCI slot I connect a PCI module made of a PCI-to-PCIe
>>>> bridge plus PCIe devices behind.
>>>>
>>>> The problem occurs when the kernel probes the PCIe devices : as
>>>> they are PCIe devices, the kernel does a PCI config read access
>>>> at offset 0x100 to check if "PCIe extended capability registers"
>>>> are accessible (see drivers/pci/probe.c, function
>>>> pci_cfg_space_size_ext()). Unfortunately the PEX8112 PCIe-to-PCI
>>>> bridge that is in the path reports an error to the CPU for this
>>>> access, and it seems there's no way to disable that on this
>>>> bridge.
>>>>
>>>> The first workaround I found was to patch
>>>> drivers/pci/host/pci-layerscape.c to have PCIE_ABSERR_SETTING set
>>>> to 0x9400 instead of 0x9401 (for PCIE_ABSERR register) to disable
>>>> error reporting. This only impacts an NXP part of the Linux
>>>> kernel code, but I'm not sure this is a good idea (however it
>>>> seems to be like that on Intel platforms where even MEM accesses
>>>> to a no-device address return FF without any error).
>>>>
>>>> I've also tried another workaround that works : patch
>>>> drivers/pci/probe.c to use bus_flags to remember if a bus is
>>>> behind a bridge without extended address capability, to avoid PCi
>>>> config read accesses at offset 0x100 in pci_cfg_space_size() /
>>>> pci_cfg_space_size_ext(). But this patch impacts the generic PCI
>>>> probe method of Linux.
>>>>
>>>> Any Idea to properly handle that issue ?
>>>>
>>> This seems like a rather unusual configuration, but I guess that
>>> if the first bridge/switch advertises its inability to support
>>> extended config space accesses, we should not be performing them
>>> on any of its subordinate buses. How does the PEX8112 advertise
>>> this limitation?
>>>
>>> That said, I wonder if it is reasonable in the first place to
>>> expect that a PCIe device works as expected passing through a
>>> legacy PCI layer like that.
>>>
>> The PEX8112 PCIe-to-PCI bridge has capability PCI_CAP_ID_EXP, but
>> has no PCI_CAP_ID_PCIX capability.  As I understand the lack of
>> PCI_CAP_ID_PCIX is advertising this limitation on the PCI side (no
>> support for PCI config offset >=0x100).
> Sounds right to me.
>
>> Also I guess in the case of a bridge having PCI_CAP_ID_PCIX, this
>> limitation would be advertised by the lack of PCI_X_STATUS_266MHZ
>> and PCI_X_STATUS_533MHZ (as done in drivers/pci/probe.c at
>> pci_cfg_space_size())
> Also sounds right.  Per the PCI-X spec, checking for PCI_X_STATUS_266MHZ
> should be enough, but it shouldn't hurt to check for either
> PCI_X_STATUS_266MHZ or PCI_X_STATUS_533MHZ.
>
>> I'm currently using the attached patch (for kernel 4.1.35-rt41 from
>> NXP Yocto BSP). It uses bus_flags to remember if a bus is behind a
>> bridge without extended address capability to avoid PCi config
>> accesses at offset >= 0x100. Thanks to this patch I now have a
>> functional system with functional PCI/PCIe devices.
> The patch seems like it's looking at the right things, but I don't
> want to build it into pci_scan_bridge_extend() because that function
> is much too complicated already.
>
> I'd rather build it into pci_cfg_space_size() or
> pci_cfg_space_size_ext() somehow.  Maybe something along these lines?
> This doesn't account for the case of a PCIe-to-PCI-X Mode 2 bridge; in
> that case, I think all 4K would be accessible on the PCI-X side.
>
> diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c
> index ac91b6fd0bcd..d8b091f0bcd1 100644
> --- a/drivers/pci/probe.c
> +++ b/drivers/pci/probe.c
> @@ -1367,7 +1367,7 @@ static bool pci_ext_cfg_is_aliased(struct pci_dev *dev)
>    * pci_cfg_space_size - Get the configuration space size of the PCI device
>    * @dev: PCI device
>    *
> - * Regular PCI devices have 256 bytes, but PCI-X 2 and PCI Express devices
> + * Regular PCI devices have 256 bytes, but PCI-X Mode 2 and PCI Express devices
>    * have 4096 bytes.  Even if the device is capable, that doesn't mean we can
>    * access it.  Maybe we don't have a way to generate extended config space
>    * accesses, or the device is behind a reverse Express bridge.  So we try
> @@ -1376,9 +1376,14 @@ static bool pci_ext_cfg_is_aliased(struct pci_dev *dev)
>    */
>   static int pci_cfg_space_size_ext(struct pci_dev *dev)
>   {
> +	struct pci_dev *bridge = pci_upstream_bridge(dev);
>   	u32 status;
>   	int pos = PCI_CFG_SPACE_SIZE;
>   
> +	if (bridge && pci_is_pcie(bridge) &&
> +	    pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE)
> +		return PCI_CFG_SPACE_SIZE;
> +
>   	if (pci_read_config_dword(dev, pos, &status) != PCIBIOS_SUCCESSFUL)
>   		return PCI_CFG_SPACE_SIZE;
>   	if (status == 0xffffffff || pci_ext_cfg_is_aliased(dev))
>
>> --- include/linux/pci.h.orig	2018-03-26 16:51:18.050000000 +0000
>> +++ include/linux/pci.h	2018-03-26 16:51:27.660000000 +0000
>> @@ -193,6 +193,7 @@
>>   enum pci_bus_flags {
>>   	PCI_BUS_FLAGS_NO_MSI   = (__force pci_bus_flags_t) 1,
>>   	PCI_BUS_FLAGS_NO_MMRBC = (__force pci_bus_flags_t) 2,
>> +	PCI_BUS_FLAGS_COMPAT_CFG_SPACE = (__force pci_bus_flags_t) 4,
>>   };
>>   
>>   /* These values come from the PCI Express Spec */
>> --- drivers/pci/probe.c.orig	2018-01-22 09:29:52.000000000 +0000
>> +++ drivers/pci/probe.c	2018-03-26 16:54:30.830000000 +0000
>> @@ -827,6 +827,28 @@
>>   			child->primary = primary;
>>   			pci_bus_insert_busn_res(child, secondary, subordinate);
>>   			child->bridge_ctl = bctl;
>> +
>> +			{
>> +				int pos;
>> +				u32 status;
>> +				bool pci_compat_cfg_space = false;
>> +
>> +				if (!pci_is_pcie(dev) || (pci_pcie_type(dev) == PCI_EXP_TYPE_PCIE_BRIDGE) || (pci_pcie_type(dev) == PCI_EXP_TYPE_PCI_BRIDGE)) {
>> +					/* for PCI/PCI bridges, or PCIe/PCI bridge in forward or reverse mode, we have to check for PCI-X capabilities */
>> +					pos = pci_find_capability(dev, PCI_CAP_ID_PCIX);
>> +					if (pos) {
>> +						pci_read_config_dword(dev, pos + PCI_X_STATUS, &status);
>> +						if (!(status & (PCI_X_STATUS_266MHZ | PCI_X_STATUS_533MHZ)))
>> +							pci_compat_cfg_space = true;
>> +					} else {
>> +						pci_compat_cfg_space = true;
>> +					}
>> +					if (pci_compat_cfg_space) {
>> +						dev_info(&dev->dev, "[%04x:%04x] Child bus limited to PCI-Compatible config space\n", dev->vendor, dev->device);
>> +						child->bus_flags |= PCI_BUS_FLAGS_COMPAT_CFG_SPACE;
>> +					}
>> +				}
>> +			}
>>   		}
>>   
>>   		cmax = pci_scan_child_bus(child);
>> @@ -1098,6 +1120,11 @@
>>   			goto fail;
>>   	}
>>   
>> +	if (dev->bus->bus_flags & PCI_BUS_FLAGS_COMPAT_CFG_SPACE) {
>> +		dev_info(&dev->dev, "[%04x:%04x] PCI-Compatible config space only due to parent bus(es)\n", dev->vendor, dev->device);
>> +		return PCI_CFG_SPACE_SIZE;
>> +	}
>> +
>>   	return pci_cfg_space_size_ext(dev);
>>   
>>    fail:
Bjorn,
If I'm right about your proposed patch to pci_cfg_space_size_ext(), *bridge is pointing to the upper device of device *dev being 
checked. I understand the purpose, but I think this fails for my config that is :

LS1043 PCIe root -> PEX8112 PCIe-to-PCI bridge -> PMC slot connector -> PCI-to-PCIe bridge -> PCIe switch (4 ports) -> 4 PCIe 
devices (one on each port)

because :
- when pci_cfg_space_size_ext() is run on the 4 PCIe devices, *bridge is the PCIe switch which is not matching 
PCI_EXP_TYPE_PCI_BRIDGE. In this case *bridge should also be checked for the parent bus of the PCIe switch, and so on.
- when pci_cfg_space_size_ext() is run for the PCI-to-PCIe bridge, *bridge is the PEX8112 that is also not matching 
PCI_EXP_TYPE_PCI_BRIDGE but PCI_EXP_TYPE_PCIE_BRIDGE. This leads to a config access at offset 0x100 to the PCI-to-PCIe bridge, so a 
crash (because of the PEX8112)

I think setting a bit in bus_flags when creating a child bus is very efficient because once set it is automatically inherited by all 
child buses and then the only thing that pci_cfg_space_size() has to do for each device is to check for this bit. Also this 
PCI_BUS_FLAGS_COMPAT_CFG_SPACE flag is actually a bus property that is compliant with the purpose of bus_flags.

I agree that pci_scan_bridge_extend() is already too complicated, so would you be okay to only add one line to it :
   pci_bus_set_compat_cfg_space(child);
and put all the code I suggested in this new function pci_bus_set_compat_cfg_space() ? (also supporting PCI-X Mode 2 devices)
Improvement : this function can return immediately if the child bus has already inherited the flag from its parent.